forum_id
stringlengths 8
20
| forum_title
stringlengths 4
171
| forum_authors
sequencelengths 0
25
| forum_abstract
stringlengths 4
4.27k
| forum_keywords
sequencelengths 1
10
| forum_pdf_url
stringlengths 38
50
| note_id
stringlengths 8
13
| note_type
stringclasses 6
values | note_created
int64 1,360B
1,736B
| note_replyto
stringlengths 8
20
| note_readers
sequencelengths 1
5
| note_signatures
sequencelengths 1
1
| note_text
stringlengths 10
16.6k
|
|---|---|---|---|---|---|---|---|---|---|---|---|---|
jRVS6C3Wia | Decoding Micromotion in Low-dimensional Latent Spaces from StyleGAN | [
"Qiucheng Wu",
"Yifan Jiang",
"Junru Wu",
"Kai Wang",
"Eric Zhang",
"Humphrey Shi",
"Zhangyang Wang",
"Shiyu Chang"
] | The disentanglement of StyleGAN latent space has paved the way for realistic and controllable image editing, but does StyleGAN know anything about temporal motion, as it was only trained on static images? To study the motion features in the latent space of StyleGAN, in this paper, we hypothesize and demonstrate that a series of meaningful, natural, and versatile small, local movements (referred to as "micromotion", such as expression, head movement, and aging effect) can be represented in low-rank spaces extracted from the latent space of a conventionally pre-trained StyleGAN-v2 model for face generation, with the guidance of proper "anchors" in the form of either short text or video clips. Starting from one target face image, with the editing direction decoded from the low-rank space, its micromotion features can be represented as simple as an affine transformation over its latent feature. Perhaps more surprisingly, such micromotion subspace, even learned from just single target face, can be painlessly transferred to other unseen face images, even those from vastly different domains (such as oil painting, cartoon, and sculpture faces). It demonstrates that the local feature geometry corresponding to one type of micromotion is aligned across different face subjects, and hence that StyleGAN-v2 is indeed ``secretly'' aware of the subject-disentangled feature variations caused by that micromotion. As an application, we present various successful examples of applying our low-dimensional micromotion subspace technique to directly and effortlessly manipulate faces. Compared with previous editing methods, our framework shows high robustness, low computational overhead, and impressive domain transferability. Our code is publicly available at https://github.com/wuqiuche/micromotion-StyleGAN. | [
"generative model",
"low-rank decomposition"
] | https://openreview.net/pdf?id=jRVS6C3Wia | l1i02dVgJW | meta_review | 1,699,323,503,770 | jRVS6C3Wia | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission31/Area_Chair_UDD4"
] | metareview: The paper makes the claim that manipulations on low rank projections of the style-GAN latent space can produce disentangled effects. Even though the effect is not fully expected theoretically, the paper produces good empirical validation and promises to release the code.
I recommend to accept the paper, conditioned on the authors publishing of the code such that all the experiments are reproducible.
recommendation: Accept (Poster)
confidence: 3: The area chair is somewhat confident |
jRVS6C3Wia | Decoding Micromotion in Low-dimensional Latent Spaces from StyleGAN | [
"Qiucheng Wu",
"Yifan Jiang",
"Junru Wu",
"Kai Wang",
"Eric Zhang",
"Humphrey Shi",
"Zhangyang Wang",
"Shiyu Chang"
] | The disentanglement of StyleGAN latent space has paved the way for realistic and controllable image editing, but does StyleGAN know anything about temporal motion, as it was only trained on static images? To study the motion features in the latent space of StyleGAN, in this paper, we hypothesize and demonstrate that a series of meaningful, natural, and versatile small, local movements (referred to as "micromotion", such as expression, head movement, and aging effect) can be represented in low-rank spaces extracted from the latent space of a conventionally pre-trained StyleGAN-v2 model for face generation, with the guidance of proper "anchors" in the form of either short text or video clips. Starting from one target face image, with the editing direction decoded from the low-rank space, its micromotion features can be represented as simple as an affine transformation over its latent feature. Perhaps more surprisingly, such micromotion subspace, even learned from just single target face, can be painlessly transferred to other unseen face images, even those from vastly different domains (such as oil painting, cartoon, and sculpture faces). It demonstrates that the local feature geometry corresponding to one type of micromotion is aligned across different face subjects, and hence that StyleGAN-v2 is indeed ``secretly'' aware of the subject-disentangled feature variations caused by that micromotion. As an application, we present various successful examples of applying our low-dimensional micromotion subspace technique to directly and effortlessly manipulate faces. Compared with previous editing methods, our framework shows high robustness, low computational overhead, and impressive domain transferability. Our code is publicly available at https://github.com/wuqiuche/micromotion-StyleGAN. | [
"generative model",
"low-rank decomposition"
] | https://openreview.net/pdf?id=jRVS6C3Wia | Xeg2BEP73s | official_review | 1,697,472,633,015 | jRVS6C3Wia | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission31/Reviewer_GYmM"
] | title: Decoding Micromotion in Low-dimensional Latent Spaces from StyleGAN
review: This paper explores the intrinsic properties of GAN hermitian spaces. The authors demonstrate that subtle movements can be represented in low-rank spaces derived from StyleGAN latent space. These micromotion features are decoded using short text or video clips as reference points. Besides, this micromotion knowledge can be transferred to different face images, even from diverse domains like paintings and sculptures. Yet, this work is great, a few questions arise:
(1). First of all, the authors have only experimented on StyleGAN-V2, and I think the generalizability of the method should be verified on more StyleGAN families, perhaps even on diffusion models.
(2). The article shows extraordinary generation results that demonstrate excellent micromotion decoupling and manipulation capabilities, but there is a little bit of blurring on the generated images in Figure.5 as well as in the upper part of the figure in Figure.7(a), Is this imperfection a natural outcome of the method, or could there be room for refinement?
(3). The experiments and supplementary experiments show some remarkably good generated results, but the presentation of the results in Figure.11 seems to be a bit blurry and distorted, could it be replaced with a clearer one?
rating: 7: Good paper, accept
confidence: 4: The reviewer is confident but not absolutely certain that the evaluation is correct |
jRVS6C3Wia | Decoding Micromotion in Low-dimensional Latent Spaces from StyleGAN | [
"Qiucheng Wu",
"Yifan Jiang",
"Junru Wu",
"Kai Wang",
"Eric Zhang",
"Humphrey Shi",
"Zhangyang Wang",
"Shiyu Chang"
] | The disentanglement of StyleGAN latent space has paved the way for realistic and controllable image editing, but does StyleGAN know anything about temporal motion, as it was only trained on static images? To study the motion features in the latent space of StyleGAN, in this paper, we hypothesize and demonstrate that a series of meaningful, natural, and versatile small, local movements (referred to as "micromotion", such as expression, head movement, and aging effect) can be represented in low-rank spaces extracted from the latent space of a conventionally pre-trained StyleGAN-v2 model for face generation, with the guidance of proper "anchors" in the form of either short text or video clips. Starting from one target face image, with the editing direction decoded from the low-rank space, its micromotion features can be represented as simple as an affine transformation over its latent feature. Perhaps more surprisingly, such micromotion subspace, even learned from just single target face, can be painlessly transferred to other unseen face images, even those from vastly different domains (such as oil painting, cartoon, and sculpture faces). It demonstrates that the local feature geometry corresponding to one type of micromotion is aligned across different face subjects, and hence that StyleGAN-v2 is indeed ``secretly'' aware of the subject-disentangled feature variations caused by that micromotion. As an application, we present various successful examples of applying our low-dimensional micromotion subspace technique to directly and effortlessly manipulate faces. Compared with previous editing methods, our framework shows high robustness, low computational overhead, and impressive domain transferability. Our code is publicly available at https://github.com/wuqiuche/micromotion-StyleGAN. | [
"generative model",
"low-rank decomposition"
] | https://openreview.net/pdf?id=jRVS6C3Wia | SIT9GkKoPi | official_review | 1,696,581,157,473 | jRVS6C3Wia | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission31/Reviewer_rDKo"
] | title: Good and interesting paper.
review: This paper proposed a method that utilizes the potential capacity of pre-trained StyleGAN to generate temporal micromotion frame sequences. Based on the hypothesis of Low-rank Micromotion Subspace, the workflow consists of three steps including Reference Anchoring which obtains a set of latent codes corresponding to the desired action performed by the same person, Robust space decomposition that utilizes robust PCA to find the edition direction, and subspace transformation that interpolate or extrapolate along the edit direction to obtain the intermediate frames.
The idea is clear and interesting. The results seem to be promising. The hypothesis is meaningful. I do recommend to accept this paper.
A question: Why the results of PCA could provide the edit direction \delta V? Could provide more detail about this? Moreover, what is the shape of \delta V?
rating: 8: Top 50% of accepted papers, clear accept
confidence: 1: The reviewer's evaluation is an educated guess |
jRVS6C3Wia | Decoding Micromotion in Low-dimensional Latent Spaces from StyleGAN | [
"Qiucheng Wu",
"Yifan Jiang",
"Junru Wu",
"Kai Wang",
"Eric Zhang",
"Humphrey Shi",
"Zhangyang Wang",
"Shiyu Chang"
] | The disentanglement of StyleGAN latent space has paved the way for realistic and controllable image editing, but does StyleGAN know anything about temporal motion, as it was only trained on static images? To study the motion features in the latent space of StyleGAN, in this paper, we hypothesize and demonstrate that a series of meaningful, natural, and versatile small, local movements (referred to as "micromotion", such as expression, head movement, and aging effect) can be represented in low-rank spaces extracted from the latent space of a conventionally pre-trained StyleGAN-v2 model for face generation, with the guidance of proper "anchors" in the form of either short text or video clips. Starting from one target face image, with the editing direction decoded from the low-rank space, its micromotion features can be represented as simple as an affine transformation over its latent feature. Perhaps more surprisingly, such micromotion subspace, even learned from just single target face, can be painlessly transferred to other unseen face images, even those from vastly different domains (such as oil painting, cartoon, and sculpture faces). It demonstrates that the local feature geometry corresponding to one type of micromotion is aligned across different face subjects, and hence that StyleGAN-v2 is indeed ``secretly'' aware of the subject-disentangled feature variations caused by that micromotion. As an application, we present various successful examples of applying our low-dimensional micromotion subspace technique to directly and effortlessly manipulate faces. Compared with previous editing methods, our framework shows high robustness, low computational overhead, and impressive domain transferability. Our code is publicly available at https://github.com/wuqiuche/micromotion-StyleGAN. | [
"generative model",
"low-rank decomposition"
] | https://openreview.net/pdf?id=jRVS6C3Wia | RwaDLqb3MC | decision | 1,700,426,755,971 | jRVS6C3Wia | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Program_Chairs"
] | decision: Accept (Oral)
comment: All reviewers and AC agreed that the paper is of high quality, this paper studies the latent space of StyleGAN and demonstrate that low rank projections of the style-GAN latent space can produce interesting disentangled effects. The work highly aligns with the theme of the conference, and presents very interesting results. The paper produces good empirical validation, and please release the code for reproducibility of the results.
The action PC chair for this paper is Qing Qu, who made the decision after carefully reading the paper as well as the comments by all reviewers and AC. The decision is agreed upon by all PC chairs.
title: Paper Decision |
jMUDBpnZPa | HARD: Hyperplane ARrangement Descent | [
"Tianjiao Ding",
"Liangzu Peng",
"Rene Vidal"
] | The problem of clustering points on a union of subspaces finds numerous applications in machine learning and computer vision, and it has been extensively studied in the past two decades. When the subspaces are low-dimensional, the problem can be formulated as a convex sparse optimization problem, for which numerous accurate, efficient and robust methods exist. When the subspaces are of high relative dimension (e.g., hyperplanes), the problem is intrinsically non-convex, and existing methods either lack theory, are computationally costly, lack robustness to outliers, or learn hyperplanes one at a time. In this paper, we propose Hyperplane ARangentment Descent (HARD), a method that robustly learns all the hyperplanes simultaneously by solving a novel non-convex non-smooth $\ell_1$ minimization problem. We provide geometric conditions under which the ground-truth hyperplane arrangement is a coordinate-wise minimizer of our objective. Furthermore, we devise efficient algorithms, and give conditions under which they converge to coordinate-wise minimizes. We provide empirical evidence that HARD surpasses state-of-the-art methods and further show an interesting experiment in clustering deep features on CIFAR-10. | [
"hyperplane clustering",
"subspace clustering",
"generalized principal component analysis"
] | https://openreview.net/pdf?id=jMUDBpnZPa | tcTBNvXDSO | meta_review | 1,699,374,932,654 | jMUDBpnZPa | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission39/Area_Chair_kq2G"
] | metareview: In this work, the authors propose Hyperplane ARrangement Descent (HARD), a novel method for clustering points on high-dimensional subspaces. HARD learns multiple hyperplanes simultaneously through a non-convex, non-smooth L1 minimization problem, outperforming existing methods and demonstrating its efficacy in clustering deep features on CIFAR-10.
This paper has received generally positive feedback from two reviewers, indicating its potential and merit. However, one reviewer expressed concerns regarding the motivation presented in the paper, as well as specific writing details. Upon thorough examination of the reviewers' comments and the authors' responses in the rebuttal, it is evident that the authors have made commendable efforts to address the concerns raised. They have provided coherent explanations and clarifications, demonstrating a willingness to enhance the paper based on the reviewers' feedback.
Considering the authors' diligent response and the potential significance of the paper, I recommend $\textbf{acceptance}$, provided that the authors make the necessary revisions as outlined in their rebuttal. These revisions should focus on strengthening the paper's motivation and addressing the specific writing concerns raised by the reviewers.
Finally, the existing convergence analysis result is relatively limited, ensuring only convergence to critical points or minimizers. It is imperative for the authors to analyze convergence to the true hyperplanes.
recommendation: Accept (Oral)
confidence: 5: The area chair is absolutely certain |
jMUDBpnZPa | HARD: Hyperplane ARrangement Descent | [
"Tianjiao Ding",
"Liangzu Peng",
"Rene Vidal"
] | The problem of clustering points on a union of subspaces finds numerous applications in machine learning and computer vision, and it has been extensively studied in the past two decades. When the subspaces are low-dimensional, the problem can be formulated as a convex sparse optimization problem, for which numerous accurate, efficient and robust methods exist. When the subspaces are of high relative dimension (e.g., hyperplanes), the problem is intrinsically non-convex, and existing methods either lack theory, are computationally costly, lack robustness to outliers, or learn hyperplanes one at a time. In this paper, we propose Hyperplane ARangentment Descent (HARD), a method that robustly learns all the hyperplanes simultaneously by solving a novel non-convex non-smooth $\ell_1$ minimization problem. We provide geometric conditions under which the ground-truth hyperplane arrangement is a coordinate-wise minimizer of our objective. Furthermore, we devise efficient algorithms, and give conditions under which they converge to coordinate-wise minimizes. We provide empirical evidence that HARD surpasses state-of-the-art methods and further show an interesting experiment in clustering deep features on CIFAR-10. | [
"hyperplane clustering",
"subspace clustering",
"generalized principal component analysis"
] | https://openreview.net/pdf?id=jMUDBpnZPa | iEBF3EQ3z1 | decision | 1,700,429,292,851 | jMUDBpnZPa | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Program_Chairs"
] | decision: Accept (Oral)
comment: All reviewers and AC agreed that the paper is of high quality. In this work, the authors propose Hyperplane ARrangement Descent (HARD), a novel method for clustering points on high-dimensional subspaces. HARD learns multiple hyperplanes simultaneously through a non-convex, non-smooth L1 minimization problem, outperforming existing methods and demonstrating its efficacy in clustering deep features on CIFAR-10. On the other hand, the existing convergence analysis result is still limited, ensuring only convergence to critical points or minimizers. It would be more impressive if the authors could show convergence to the true hyperplanes.
The action PC chair for this paper is Qing Qu, who made the decision after carefully reading the paper as well as the comments by all reviewers and AC. The decision is agreed upon by all PC chairs.
title: Paper Decision |
jMUDBpnZPa | HARD: Hyperplane ARrangement Descent | [
"Tianjiao Ding",
"Liangzu Peng",
"Rene Vidal"
] | The problem of clustering points on a union of subspaces finds numerous applications in machine learning and computer vision, and it has been extensively studied in the past two decades. When the subspaces are low-dimensional, the problem can be formulated as a convex sparse optimization problem, for which numerous accurate, efficient and robust methods exist. When the subspaces are of high relative dimension (e.g., hyperplanes), the problem is intrinsically non-convex, and existing methods either lack theory, are computationally costly, lack robustness to outliers, or learn hyperplanes one at a time. In this paper, we propose Hyperplane ARangentment Descent (HARD), a method that robustly learns all the hyperplanes simultaneously by solving a novel non-convex non-smooth $\ell_1$ minimization problem. We provide geometric conditions under which the ground-truth hyperplane arrangement is a coordinate-wise minimizer of our objective. Furthermore, we devise efficient algorithms, and give conditions under which they converge to coordinate-wise minimizes. We provide empirical evidence that HARD surpasses state-of-the-art methods and further show an interesting experiment in clustering deep features on CIFAR-10. | [
"hyperplane clustering",
"subspace clustering",
"generalized principal component analysis"
] | https://openreview.net/pdf?id=jMUDBpnZPa | LHduErZsPg | official_review | 1,696,748,474,264 | jMUDBpnZPa | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission39/Reviewer_qtDd"
] | title: The theory is interesting and educational as it connects GPCA and DPCP formalisms. Though, the motivation is weak, the writing and the structure of the paper can be greatly improved, and detailed discussions are missing.
review: This paper studies the problem on subspace clustering. The main emphasis of the work are (1) hyperplane clustering, (2) using cost functions that enable robustness to outlier data, namely, L1 and Huber and their variations, (3) providing theoretical guarantees on various notions of the optimality of estimated hyperplanes --- under strong conditions on data points and outliers. The experiments in the main text include numerical results on small-scale synthetic data and CIFAR-10 features.
####################################################################
My evaluation:
This paper introduces interesting theoretical results that combine GPCA formulation of hyperspheres with robust L1 clustering cost (inspired by DPCP) and prove certain optimalities for estimated hyperspaces under strong geometric conditions. However, this does not qualify to be accepted due to the following issues: (a) The motivation of using "only" hyperspaces is not fully explained. What happens when we have subspaces of codimensions >2? (b) The writing the paper is poor : unscientific word and claims, the structure, explanation of the assumptions 5.1 is insufficient, ... (c) The experiments includes small-scale data. What happens when we have no outliers or we have different variances for random additive noises? You can should move some of the "no-outlier" results in the appendix to the main text and investigate how this method compares in estimating the correct subspace with other methods (in the no-outlier scenario).
P.S. I assume that the proposed Theorems are correct as I did not read proofs in the Appendix.
####################################################################
I will summarize my main questions, comments, and suggestions as follows:
(1) Regarding the "Geometric Quantities": Suppose datapoints belong to subspace of codimension 2 (a subspace of the supposed hyperplane). Then, it is easy to show that c_{in, k , min} is 0 for all k \in [K]. In this case, we are not guaranteed to have {b^{*}_k} as coordinate-wise minimizers of (GPCA-l1) --- according to Theorem 4.3. The authors should explain what happens in this situation --- which is quite common in practice.
(2) Related to my comment (1), the proposed method formalizes the problem of subspace clustering using orthogonal complement representation of the hyperspaces, that is, vectors { b_k }. Therefore, once these vectors estimated, the algorithm returns the estimated hyperspaces. One can easily think of a situation where we want to estimate high-dimensional subspaces with codimensions > 1. What would be the appropriate modification of this algorithm to operate on those cases? One may think that you have to resort to estimating one-dimension at a time in the orthogonal complement space --- which would be greedy. It is important to clearly explain your approach from this viewpoint so that the reader can easily compare this algorithm with other methods.
(3) Assumption 5.1 might be standard but not realistic at all. Either the authors should provide their insight on why (or on which class of point sets) this assumption is true or clearly explain that this assumption puts a strong set of constraints on both inlier and outlier point sets -- which maybe even impossible to satisfy. Regarding (A3), please either prove the claim that for random points the eigenvectors of the weighted sum are distinct (with probability 1) or cite a reference.
(4) (page 2, line 43) “KH-DPCP [27] integrates DPCP into the K-Hyperplanes framework. “ + (page 3, line 102) “Second, the underlying theory of why such an integration works well has thus far remained, to our knowledge, obscure. “
The algorithm KH-DPCP is not discussed in the main text. Given the fact that its objective is the most comparable to this problem (robust hyperplane clustering), it is beneficial to dedicate a detailed remark on this algorithm.
In what follows, I give comments on issues that appear frequently in the paper (not limited to these examples).
(5) The specific choice of notations may seem irrelevant. But from the perspective of readers, authors should try to use “intuitive” notations. Two examples (among many): (a) In equation (GPCA-l2), indices in summation goes from j=1 to N. why not n=1 to N? (b) The definition of w^{t}_{j,k} in Algorithm 1: HARD-l is "similar" to that of d^{k}_j in subsection 4.1. (c) A good notation can simplify Equation (2) and make it more intuitive, namely, I - bb^{T} is an orthogonal projection matrix ...
(6) (page 1, line 24) “However, for subspaces of high relative dimensions (relative to D), sparsity and low-rankness break down, and so do these methods.” The word relative is repetitive. The sentence is broken.
(7) (page 1, line 36) “It is very simple and intuitive, but it is inaccurate and not robust to outliers, and it has limited theoretical guarantees (e.g., of convergence to true hyperplanes). “
Please do not use words that do not convey useful information, like “very”. Please also do not construct run on sentences ( … but … and … and … ). This hinders the readability of your paper.
(8) (page 2, line 48) “we blend the GPCA and DPCP philosophies”
I’m not sure if “philosophies” is an appropriate word here.
(9) (page 2, line 44) “In doing so, it inherits the one-shot ability of K-Hyperplanes and the robustness of DPCP to a certain extent. But it also compromises accuracy and comes with no theoretical guarantees”
“to a certain extent” does not make sense to me. Could you please explain this algorithm, and compare it with your proposed algorithms?
(10) (page 3, line 92) “The first idea of [27] has some … ” Please do not use a numerical reference as an object in a sentence. You can use the authors’ names instead.
(11) (page 4, line 127) “How can we solve the harder problem (GPCA-l1) efficiently?“ I am not sure in what sense the problem is hard. Maybe instead you can emphasize on the fact that the l1 objective is nonsmooth?
(12) (page 4, line 132) Whether the true normal vectors are a global minimizer of (GPCA-l1)? I do not understand the meaning of this sentence.
(13) (page 7, line 229) "Intuitively, (A2) is more likely than (A1) to have a unique global minimizer because h is further locally quadratic (strongly convex)." The term likely implies the existence of an underlying probability distribution. If there in none, this is an incorrect statement.
rating: 5: Marginally below acceptance threshold
confidence: 4: The reviewer is confident but not absolutely certain that the evaluation is correct |
jMUDBpnZPa | HARD: Hyperplane ARrangement Descent | [
"Tianjiao Ding",
"Liangzu Peng",
"Rene Vidal"
] | The problem of clustering points on a union of subspaces finds numerous applications in machine learning and computer vision, and it has been extensively studied in the past two decades. When the subspaces are low-dimensional, the problem can be formulated as a convex sparse optimization problem, for which numerous accurate, efficient and robust methods exist. When the subspaces are of high relative dimension (e.g., hyperplanes), the problem is intrinsically non-convex, and existing methods either lack theory, are computationally costly, lack robustness to outliers, or learn hyperplanes one at a time. In this paper, we propose Hyperplane ARangentment Descent (HARD), a method that robustly learns all the hyperplanes simultaneously by solving a novel non-convex non-smooth $\ell_1$ minimization problem. We provide geometric conditions under which the ground-truth hyperplane arrangement is a coordinate-wise minimizer of our objective. Furthermore, we devise efficient algorithms, and give conditions under which they converge to coordinate-wise minimizes. We provide empirical evidence that HARD surpasses state-of-the-art methods and further show an interesting experiment in clustering deep features on CIFAR-10. | [
"hyperplane clustering",
"subspace clustering",
"generalized principal component analysis"
] | https://openreview.net/pdf?id=jMUDBpnZPa | E855Lj95Aa | official_review | 1,696,447,909,377 | jMUDBpnZPa | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission39/Reviewer_vpD9"
] | title: Solid work
review: This paper proposed a new algorithm for Hyperplane (2D subspace) clustering algorithm called hyperplane arrangement descent (HARD) that is both efficient and robust to outliers. This new family of algorithms solve a robust version of the Generalized Principal Component Analysis (GPCA) objective that uses a L1 norm for improved robustness against outliers, this new objective is called GPCA-l1.
This proposed GPCA-l1 objective can be optimized by a coordinate-wise algorithm called HARD-l1, however, this algorithm contains an inner optimization step that itself is cumbersome to optimize. Therefore the authors proposed to relax the inner optimization problem to a smoothed upper bound that can be efficiently solvable by SVD. The resulting algorithm is called HARD-l1+. The authors then study a smoother version of the GPCA-l1 that is called GPCA-Huber. Similar to HARD-l1, the authors relaxed the inner optimization of HARD-Huber to obtain a faster algorithm HARD-Huber+.
The authors move on to prove that the ground truth hyperplane arrangement minimizes the GPCA-l1 objective, and that the HARD-l1, HARD-Huber and HARD-Huber+ algorithm can successfully find critical points of GPCA-l1, GPCA-Huber and GPCA-Huber+ respectively at convergence. However, the authors find it difficult to prove that the practical HARD-Huber+ algorithm converges to a critical point.
Lastly, the authors used synthetic datasets to show that HARD-l1+ and HARD-Huber+ achieves better clustering accuracy as well as significantly superior runtime compared to other subspace or hyperplane clustering algorithms. The new algorithms also outperform previous algorithms on a more difficult and realistic clustering problem involving latent embedding learned by a self-supervised learning algorithm on CIFAR-10.
Overall, the paper is very well written and clear, the logic is easy to follow. Although the practical algorithms don't have theoretical guarantees, their benefit is clearly demonstrated. I look forward to seeing more applications of those algorithms in the ML community.
rating: 8: Top 50% of accepted papers, clear accept
confidence: 3: The reviewer is fairly confident that the evaluation is correct |
jMUDBpnZPa | HARD: Hyperplane ARrangement Descent | [
"Tianjiao Ding",
"Liangzu Peng",
"Rene Vidal"
] | The problem of clustering points on a union of subspaces finds numerous applications in machine learning and computer vision, and it has been extensively studied in the past two decades. When the subspaces are low-dimensional, the problem can be formulated as a convex sparse optimization problem, for which numerous accurate, efficient and robust methods exist. When the subspaces are of high relative dimension (e.g., hyperplanes), the problem is intrinsically non-convex, and existing methods either lack theory, are computationally costly, lack robustness to outliers, or learn hyperplanes one at a time. In this paper, we propose Hyperplane ARangentment Descent (HARD), a method that robustly learns all the hyperplanes simultaneously by solving a novel non-convex non-smooth $\ell_1$ minimization problem. We provide geometric conditions under which the ground-truth hyperplane arrangement is a coordinate-wise minimizer of our objective. Furthermore, we devise efficient algorithms, and give conditions under which they converge to coordinate-wise minimizes. We provide empirical evidence that HARD surpasses state-of-the-art methods and further show an interesting experiment in clustering deep features on CIFAR-10. | [
"hyperplane clustering",
"subspace clustering",
"generalized principal component analysis"
] | https://openreview.net/pdf?id=jMUDBpnZPa | BvtWxfGyeS | official_review | 1,696,205,568,388 | jMUDBpnZPa | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission39/Reviewer_m9My"
] | title: Report on HARD: Hyperplane ARrangement Descent
review: In this paper, the authors address the hyperplane clustering problem in the presence of outliers. They introduce the Hyperplane Arrangement Descent (HARD) algorithms and provide a comprehensive theoretical analysis. In my view, this work represents a significant accomplishment.
Comments:
1. It would be beneficial to include specific details regarding the computation of $b_k^{(t+1)}$ and elaborate on their computational complexity. Additionally, a comparative analysis of the computational complexity of your proposed method with other approaches based on different loss functions would enhance the paper's clarity.
2. Figure 2 needs further clarification. Please define the x-axis to ensure its meaning is clear to readers. Furthermore, could you explain why the running time of the red lines is higher at $K=4$ than at $K=5$ in Figure 2(c)?
rating: 7: Good paper, accept
confidence: 5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature |
iW26qcPlui | FIXED: Frustratingly Easy Domain Generalization with Mixup | [
"Wang Lu",
"Jindong Wang",
"Han Yu",
"Lei Huang",
"Xiang Zhang",
"Yiqiang Chen",
"Xing Xie"
] | Domain generalization (DG) aims to learn a generalizable model from multiple training domains such that it can perform well on unseen target domains. A popular strategy is to augment training data to benefit generalization through methods such as Mixup [1]. While the vanilla Mixup can be directly applied, theoretical and empirical investigations uncover several shortcomings that limit its performance. Firstly, Mixup cannot effectively identify the domain and class information that can be used for learning invariant representations. Secondly, Mixup may introduce synthetic noisy data points via random interpolation, which lowers its discrimination capability. Based on the analysis, we propose a simple yet effective enhancement for Mixup-based DG, namely domain-invariant Feature mIXup (FIX). It learns domain-invariant representations for Mixup. To further enhance discrimination, we leverage existing techniques to enlarge margins among classes to further propose the domain-invariant Feature MIXup with Enhanced Discrimination (FIXED) approach. We present theoretical insights about guarantees on its effectiveness. Extensive experiments on seven public datasets across two modalities including image classification (Digits-DG, PACS, Office-Home) and time series (DSADS, PAMAP2, UCI-HAR, and USC-HAD) demonstrate that our approach significantly outperforms nine state-of-the-art related methods, beating the best performing baseline by 6.5% on average in terms of test accuracy. The code is available at https:// github.com/jindongwang/transferlearning/tree/master/code/deep/fixed. | [
"Domain generalization",
"Data Augmentation",
"Out-of-distribution generalization"
] | https://openreview.net/pdf?id=iW26qcPlui | r6DyuGz5Iy | official_review | 1,696,725,220,286 | iW26qcPlui | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission43/Reviewer_nEJf"
] | title: Review for FIXED: Frustratingly Easy Domain Generalization with Mixup
review: Summary Of The Paper
This paper studies domain adaption by performing augmentations. The authors introduce a method, domain-invariant feature mixup, which is an enhancement of Mixup. This paper also introduces a margin loss to better discriminate among each class.
Main Review
- Strength
1) The theoretical validation are shown to proof its effectiveness. And the analysis on two aspects: 1) distribution coverage and 2) inter-class distance, is informative.
2) The proposed method is extensively compared against other methods on several dataset including image classification and time series, and is showing the effectiveness on moment retrieval and highlight detection tasks. And it is consistently better than other methods.
- Weakness
1) With extensive experiments and analysis, the paper has demonstrated its proposed strategy on several dataset, however, the intuition of the proposed method is not thoroughly discussed, and the novelty of this work is somewhat not very strong because the paper is adapting existing strategies (feature mixup, margin loss) together into the learning scheme.
2) In order to demonstrate the effectiveness of the proposed method, larger scale image classification dataset based on natural images, such as ImagNet must be explored, at least the author should try ImageNet-100 (https://github.com/danielchyeh/ImageNet-100-Pytorch) data, which is a subset of ImageNet. I think the proposed method may not be practical without such larger scale dataset.
Summary Of The Review
Overall, without discussing the novelty of the proposed method in details, I am not sure if the novelty that authors mention in the paper is reliable. Also, I think we need to see valid elaboration and the intuition of the proposed method, and scale up to larger dataset on image classification task. Combined with the weaknesses I mentioned above, I vote for 5. I would like to see authors response to consider raising the rating.
rating: 5: Marginally below acceptance threshold
confidence: 4: The reviewer is confident but not absolutely certain that the evaluation is correct |
iW26qcPlui | FIXED: Frustratingly Easy Domain Generalization with Mixup | [
"Wang Lu",
"Jindong Wang",
"Han Yu",
"Lei Huang",
"Xiang Zhang",
"Yiqiang Chen",
"Xing Xie"
] | Domain generalization (DG) aims to learn a generalizable model from multiple training domains such that it can perform well on unseen target domains. A popular strategy is to augment training data to benefit generalization through methods such as Mixup [1]. While the vanilla Mixup can be directly applied, theoretical and empirical investigations uncover several shortcomings that limit its performance. Firstly, Mixup cannot effectively identify the domain and class information that can be used for learning invariant representations. Secondly, Mixup may introduce synthetic noisy data points via random interpolation, which lowers its discrimination capability. Based on the analysis, we propose a simple yet effective enhancement for Mixup-based DG, namely domain-invariant Feature mIXup (FIX). It learns domain-invariant representations for Mixup. To further enhance discrimination, we leverage existing techniques to enlarge margins among classes to further propose the domain-invariant Feature MIXup with Enhanced Discrimination (FIXED) approach. We present theoretical insights about guarantees on its effectiveness. Extensive experiments on seven public datasets across two modalities including image classification (Digits-DG, PACS, Office-Home) and time series (DSADS, PAMAP2, UCI-HAR, and USC-HAD) demonstrate that our approach significantly outperforms nine state-of-the-art related methods, beating the best performing baseline by 6.5% on average in terms of test accuracy. The code is available at https:// github.com/jindongwang/transferlearning/tree/master/code/deep/fixed. | [
"Domain generalization",
"Data Augmentation",
"Out-of-distribution generalization"
] | https://openreview.net/pdf?id=iW26qcPlui | ZaccBtfk5b | meta_review | 1,699,838,935,391 | iW26qcPlui | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission43/Area_Chair_ZB9b"
] | metareview: This paper proposes a modified version of Mixup for improved domain generalization. The paper provides both theoretical insights and empirical evaluation showing improvements. The reviewers generally acknowledge the quality and significance of the proposed method as well as the extensive empirical evaluation. On the negative side, 3 out of 4 reviewers point out the lack of clarity in the description of the intuition/motivations/mechanism of the proposed method. The authors should work on improving this clarity in the final version.
recommendation: Accept (Poster)
confidence: 4: The area chair is confident but not absolutely certain |
iW26qcPlui | FIXED: Frustratingly Easy Domain Generalization with Mixup | [
"Wang Lu",
"Jindong Wang",
"Han Yu",
"Lei Huang",
"Xiang Zhang",
"Yiqiang Chen",
"Xing Xie"
] | Domain generalization (DG) aims to learn a generalizable model from multiple training domains such that it can perform well on unseen target domains. A popular strategy is to augment training data to benefit generalization through methods such as Mixup [1]. While the vanilla Mixup can be directly applied, theoretical and empirical investigations uncover several shortcomings that limit its performance. Firstly, Mixup cannot effectively identify the domain and class information that can be used for learning invariant representations. Secondly, Mixup may introduce synthetic noisy data points via random interpolation, which lowers its discrimination capability. Based on the analysis, we propose a simple yet effective enhancement for Mixup-based DG, namely domain-invariant Feature mIXup (FIX). It learns domain-invariant representations for Mixup. To further enhance discrimination, we leverage existing techniques to enlarge margins among classes to further propose the domain-invariant Feature MIXup with Enhanced Discrimination (FIXED) approach. We present theoretical insights about guarantees on its effectiveness. Extensive experiments on seven public datasets across two modalities including image classification (Digits-DG, PACS, Office-Home) and time series (DSADS, PAMAP2, UCI-HAR, and USC-HAD) demonstrate that our approach significantly outperforms nine state-of-the-art related methods, beating the best performing baseline by 6.5% on average in terms of test accuracy. The code is available at https:// github.com/jindongwang/transferlearning/tree/master/code/deep/fixed. | [
"Domain generalization",
"Data Augmentation",
"Out-of-distribution generalization"
] | https://openreview.net/pdf?id=iW26qcPlui | R64B5rktef | official_review | 1,696,614,595,360 | iW26qcPlui | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission43/Reviewer_Pg71"
] | title: Review of Submission 43
review: **Overall Evaluation**
This paper approaches domain generalization through the lens of the Mixup technique, highlighting current challenges including the entanglement of domain and class information and the potential pitfalls of introducing noisy data. Its novelty is underscored by demonstrating the adverse effects of synthetic data generation, especially in terms of domain and class entanglement, not only in conceptual illustration but also within real-world datasets.
*Pros*
1. The negative effect of entanglement on domain and class information is well proposed, and its remedy is straightforward.
2. Analytical evaluation of distribution coverage and inter-class distance further strengthens the motivation of FIXED.
3. Experiments across two modalities are well-conducted and show convincing results.
*Cons*
1. While the primary focus of this paper is on differentiating between domain and class information, offering a deeper discussion on how their relationship (i.e., entanglement) can occasionally serve beneficial roles in real-world scenarios—such as instances where a class originates from distinct domains—would enhance the paper's motivation.
2. Sensitivity analysis could be more comprehensive. The impact of adversarial learning weight and the required distance to boundaries may be intertwined. Presenting their joint effects using a 2D heatmap or 3D plot could offer deeper insights.
3. The visual presentation in this paper, including figures and tables, needs improvement. The space for Figure 4 and Table 3 should be expanded for clarity. Figure 5, in particular, appears too small and could benefit from resizing. However, these can be easily addressed in their final version.
rating: 8: Top 50% of accepted papers, clear accept
confidence: 4: The reviewer is confident but not absolutely certain that the evaluation is correct |
iW26qcPlui | FIXED: Frustratingly Easy Domain Generalization with Mixup | [
"Wang Lu",
"Jindong Wang",
"Han Yu",
"Lei Huang",
"Xiang Zhang",
"Yiqiang Chen",
"Xing Xie"
] | Domain generalization (DG) aims to learn a generalizable model from multiple training domains such that it can perform well on unseen target domains. A popular strategy is to augment training data to benefit generalization through methods such as Mixup [1]. While the vanilla Mixup can be directly applied, theoretical and empirical investigations uncover several shortcomings that limit its performance. Firstly, Mixup cannot effectively identify the domain and class information that can be used for learning invariant representations. Secondly, Mixup may introduce synthetic noisy data points via random interpolation, which lowers its discrimination capability. Based on the analysis, we propose a simple yet effective enhancement for Mixup-based DG, namely domain-invariant Feature mIXup (FIX). It learns domain-invariant representations for Mixup. To further enhance discrimination, we leverage existing techniques to enlarge margins among classes to further propose the domain-invariant Feature MIXup with Enhanced Discrimination (FIXED) approach. We present theoretical insights about guarantees on its effectiveness. Extensive experiments on seven public datasets across two modalities including image classification (Digits-DG, PACS, Office-Home) and time series (DSADS, PAMAP2, UCI-HAR, and USC-HAD) demonstrate that our approach significantly outperforms nine state-of-the-art related methods, beating the best performing baseline by 6.5% on average in terms of test accuracy. The code is available at https:// github.com/jindongwang/transferlearning/tree/master/code/deep/fixed. | [
"Domain generalization",
"Data Augmentation",
"Out-of-distribution generalization"
] | https://openreview.net/pdf?id=iW26qcPlui | Prmd9fNcbq | official_review | 1,697,404,010,276 | iW26qcPlui | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission43/Reviewer_uvaF"
] | title: Official Review of Submission43 by Reviewer uvaF
review: ### Contribution:
The submission introduces FIXED, a nuanced approach aiming to refine domain generalization through a modified Mixup process. The authors articulate theoretical underpinnings for their strategy and validate their claims with experiments, suggesting broader applicability in classification tasks.
### Strengths:
1. FIXED presents a thoughtful attempt to navigate the complexities of domain-invariant representation, potentially enhancing model robustness. The approach seems promising within the scope of the presented experiments.
2. The submission delves into a detailed theoretical discourse, shedding light on the intricacies of domain generalization and the proposed method's potential advantages.
3. The empirical work is commendable, with a rigorous experimental setup and a comprehensive analysis that lends credibility to the proposed method's efficacy.
### Weaknesses:
1. The narrative detailing FIXED's operational mechanism, especially its handling of domain and class information, lacks clarity. This obscurity might hinder readers' understanding and raise questions about the method's adaptability.
2. The dense theoretical exposition in Sec. 4, while insightful, poses accessibility issues. A more approachable presentation of these complex concepts would likely benefit a wider audience.
3. The submission would gain from a more nuanced discussion on FIXED's limitations, computational demands, and its behavior under varied conditions, which remains unexplored.
4. On a minor note, the formatting issues require attention to enhance the presentation quality. Specifically, elements like Figure 4, Table 2, and Table 3 are too close to the text margins, compromising readability. Additionally, the absence of error bars in the experimental results is a concern. Incorporating error bars would significantly strengthen the solidity of the findings by providing a clearer depiction of data variability. I kindly suggest these refinements to ensure a more polished and authoritative paper.
### Broader Impact Concerns:
The submission does not directly engage with the broader implications of the proposed method. It would be beneficial for the authors to speculate on both the positive and negative ramifications of their work in real-world contexts, ensuring a holistic understanding.
### Conclusion:
This paper makes a tentative step forward in the realm of domain generalization, offering a potentially valuable method with FIXED. While the theoretical and empirical aspects are generally well-executed, there are areas where the clarity of communication and depth of analysis could be enhanced. The paper somewhat meets the conference's criteria, suggesting that, with further refinement in the areas highlighted, it could resonate with the interests of a segment of the CPAL audience. Therefore, I lean towards borderline acceptance, contingent on the authors' willingness to address the identified concerns.
rating: 6: Marginally above acceptance threshold
confidence: 3: The reviewer is fairly confident that the evaluation is correct |
iW26qcPlui | FIXED: Frustratingly Easy Domain Generalization with Mixup | [
"Wang Lu",
"Jindong Wang",
"Han Yu",
"Lei Huang",
"Xiang Zhang",
"Yiqiang Chen",
"Xing Xie"
] | Domain generalization (DG) aims to learn a generalizable model from multiple training domains such that it can perform well on unseen target domains. A popular strategy is to augment training data to benefit generalization through methods such as Mixup [1]. While the vanilla Mixup can be directly applied, theoretical and empirical investigations uncover several shortcomings that limit its performance. Firstly, Mixup cannot effectively identify the domain and class information that can be used for learning invariant representations. Secondly, Mixup may introduce synthetic noisy data points via random interpolation, which lowers its discrimination capability. Based on the analysis, we propose a simple yet effective enhancement for Mixup-based DG, namely domain-invariant Feature mIXup (FIX). It learns domain-invariant representations for Mixup. To further enhance discrimination, we leverage existing techniques to enlarge margins among classes to further propose the domain-invariant Feature MIXup with Enhanced Discrimination (FIXED) approach. We present theoretical insights about guarantees on its effectiveness. Extensive experiments on seven public datasets across two modalities including image classification (Digits-DG, PACS, Office-Home) and time series (DSADS, PAMAP2, UCI-HAR, and USC-HAD) demonstrate that our approach significantly outperforms nine state-of-the-art related methods, beating the best performing baseline by 6.5% on average in terms of test accuracy. The code is available at https:// github.com/jindongwang/transferlearning/tree/master/code/deep/fixed. | [
"Domain generalization",
"Data Augmentation",
"Out-of-distribution generalization"
] | https://openreview.net/pdf?id=iW26qcPlui | IZLVVNgGnL | official_review | 1,697,473,263,232 | iW26qcPlui | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission43/Reviewer_gZbi"
] | title: solid experiments; questions on the motivation
review: Summary:
This work introduces a modified Mixup technique to address domain generation. Specifically, the mixup is performed over domain-invariant features (generated by DANN), and a large margin loss is introduced to complement the Mixup loss. Thorough experiments are conducted, together with theoretical insights.
Advantages:
• The approach is simple and easy to implement.
• Experimental evaluations are rather complete and showcase the method's efficacy.
Downsides:
• Despite the empirical improvement, the motivations are unconvincing to me. Specifically, the arguments made for Figure 1 all hinge on the fact that the toy example is very low dimensional. I am skeptical that the mixed samples will likely overlap with existing clusters. Also, I am unsure why the domain information becomes an issue -- the same argument could be made for vanilla classification tasks in one domain.
• The approach, in essence, combines several pieces of existing techniques, including domain-invariant features, manifold mixup, and margin loss, together with their hyper-parameters. So, I am unsurprised to witness the empirical gains, given the additional degrees of freedom for hyperparameter tuning.
• I would appreciate a comparison with another mix-up-based approach for distribution shift [1].
[1] https://arxiv.org/pdf/2201.00299.pdf
rating: 5: Marginally below acceptance threshold
confidence: 3: The reviewer is fairly confident that the evaluation is correct |
iW26qcPlui | FIXED: Frustratingly Easy Domain Generalization with Mixup | [
"Wang Lu",
"Jindong Wang",
"Han Yu",
"Lei Huang",
"Xiang Zhang",
"Yiqiang Chen",
"Xing Xie"
] | Domain generalization (DG) aims to learn a generalizable model from multiple training domains such that it can perform well on unseen target domains. A popular strategy is to augment training data to benefit generalization through methods such as Mixup [1]. While the vanilla Mixup can be directly applied, theoretical and empirical investigations uncover several shortcomings that limit its performance. Firstly, Mixup cannot effectively identify the domain and class information that can be used for learning invariant representations. Secondly, Mixup may introduce synthetic noisy data points via random interpolation, which lowers its discrimination capability. Based on the analysis, we propose a simple yet effective enhancement for Mixup-based DG, namely domain-invariant Feature mIXup (FIX). It learns domain-invariant representations for Mixup. To further enhance discrimination, we leverage existing techniques to enlarge margins among classes to further propose the domain-invariant Feature MIXup with Enhanced Discrimination (FIXED) approach. We present theoretical insights about guarantees on its effectiveness. Extensive experiments on seven public datasets across two modalities including image classification (Digits-DG, PACS, Office-Home) and time series (DSADS, PAMAP2, UCI-HAR, and USC-HAD) demonstrate that our approach significantly outperforms nine state-of-the-art related methods, beating the best performing baseline by 6.5% on average in terms of test accuracy. The code is available at https:// github.com/jindongwang/transferlearning/tree/master/code/deep/fixed. | [
"Domain generalization",
"Data Augmentation",
"Out-of-distribution generalization"
] | https://openreview.net/pdf?id=iW26qcPlui | ALpsQv08G5 | decision | 1,700,423,011,925 | iW26qcPlui | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Program_Chairs"
] | decision: Accept (Oral)
comment: This work proposes a simple enhancement for Mixup-based domain generalization, and demonstrates its effectiveness through extensive numerical experiments, which will be a good addition to the conference. The authors should improve the writing clarity for the final paper.
The action PC chair for this paper is Yuejie Chi, who made the decision after carefully reading the paper as well as the comments by all reviewers and AC. The decision is agreed by all PC chairs.
title: Paper Decision |
hJd66ZzXEZ | Domain Generalization via Nuclear Norm Regularization | [
"Zhenmei Shi",
"Yifei Ming",
"Ying Fan",
"Frederic Sala",
"Yingyu Liang"
] | The ability to generalize to unseen domains is crucial for machine learning systems deployed in the real world, especially when we only have data from limited training domains. In this paper, we propose a simple and effective regularization method based on the nuclear norm of the learned features for domain generalization. Intuitively, the proposed regularizer mitigates the impacts of environmental features and encourages learning domain-invariant features. Theoretically, we provide insights into why nuclear norm regularization is more effective compared to ERM and alternative regularization methods. Empirically, we conduct extensive experiments on both synthetic and real datasets. We show nuclear norm regularization achieves strong performance compared to baselines in a wide range of domain generalization tasks. Moreover, our regularizer is broadly applicable with various methods such as ERM and SWAD with consistently improved performance, e.g., 1.7% and 0.9% test accuracy improvements respectively on the DomainBed benchmark. | [
"Domain Generalization",
"Nuclear Norm",
"Deep Learning"
] | https://openreview.net/pdf?id=hJd66ZzXEZ | pFvZnZryXf | decision | 1,700,420,646,321 | hJd66ZzXEZ | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Program_Chairs"
] | decision: Accept (Oral)
comment: The paper tackles the important problem domain generalization using a novel nuclear norm regularization, with both theoretical and empirical justifications. The writing of the paper is clear and easy-to-follow.
The action PC chair for this paper is Yuejie Chi, who made the decision after carefully reading the paper as well as the comments by all reviewers and AC. The decision is agreed by all PC chairs.
title: Paper Decision |
hJd66ZzXEZ | Domain Generalization via Nuclear Norm Regularization | [
"Zhenmei Shi",
"Yifei Ming",
"Ying Fan",
"Frederic Sala",
"Yingyu Liang"
] | The ability to generalize to unseen domains is crucial for machine learning systems deployed in the real world, especially when we only have data from limited training domains. In this paper, we propose a simple and effective regularization method based on the nuclear norm of the learned features for domain generalization. Intuitively, the proposed regularizer mitigates the impacts of environmental features and encourages learning domain-invariant features. Theoretically, we provide insights into why nuclear norm regularization is more effective compared to ERM and alternative regularization methods. Empirically, we conduct extensive experiments on both synthetic and real datasets. We show nuclear norm regularization achieves strong performance compared to baselines in a wide range of domain generalization tasks. Moreover, our regularizer is broadly applicable with various methods such as ERM and SWAD with consistently improved performance, e.g., 1.7% and 0.9% test accuracy improvements respectively on the DomainBed benchmark. | [
"Domain Generalization",
"Nuclear Norm",
"Deep Learning"
] | https://openreview.net/pdf?id=hJd66ZzXEZ | gPnpdQG0Az | official_review | 1,696,634,098,788 | hJd66ZzXEZ | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission5/Reviewer_odop"
] | title: Solid paper, may need additional empirical verification.
review: The paper proposes a simple yet effective method that leverages nuclear norm regularization to address the neural network's tendency to capture spurious correlations between labels and images instead of domain-invariant features. The authors provide a comprehensive set of experiments on synthetic and real datasets to showcase the effectiveness of their proposed method. Additionally, the inclusion of theoretical support in simpler settings improves the credibility of their approach.
## Pros and Cons
### Pros
The paper is clearly written and well-presented, all notations, and definitions are clearly conveyed, making it an enjoyable reading experience. Furthermore, most of the claims and conjectures mentioned in the paper are backed up using either empirical or theoretical results, enhancing the overall credibility of the research.
### Cons
The reviewer finds no major flaws in the paper. However, one aspect that raises some questions is the motivation presented in the paper: 'our main hypothesis is that environmental features have a lower correlation with the label than the invariant features.' While the theoretical analysis in Section 4 provides some support for this hypothesis, there is a lack of empirical evidence to reinforce this claim. Given that the setting in Section 4 is a simplistic setting, the reviewer believes that the intuition is not firmly established. Therefore, perhaps adding some experiments to empirically validate the motivation is useful. One starting point could be using the synthetic data introduced in Section 3.1 to visualize the magnitude of the matrix $A$ associated with environmental features and domain-invariant features, respectively, and see if the results meet the claims in the intuition.
## Questions
Q1: the notion $w$ is repeatedly used In line 259 when defining features and in line 284/285 to refer to the optimization variable (the entries of the classifier vector), which seems a bit confusing. The reviewer can understand that $w$ represents the strength of a corresponding feature, but why does it appear twice here?
Q2: In Proposition 3, the conclusion from the authors is that the OOD accuracy of ERM objective is "much worse than random guessing", while clearly, this is not the case in reality as we can observe from all the empirical results in the manuscript. Can the authors say more about the discrepancy between the theory and practice to explain this gap?
Q3: since the proposed method involves adding additional regularization on loss functions, the regularization strength becomes a tuning parameter. Do the authors conduct hyperparameter tuning for all tasks to find the suitable parameter?
rating: 7: Good paper, accept
confidence: 4: The reviewer is confident but not absolutely certain that the evaluation is correct |
hJd66ZzXEZ | Domain Generalization via Nuclear Norm Regularization | [
"Zhenmei Shi",
"Yifei Ming",
"Ying Fan",
"Frederic Sala",
"Yingyu Liang"
] | The ability to generalize to unseen domains is crucial for machine learning systems deployed in the real world, especially when we only have data from limited training domains. In this paper, we propose a simple and effective regularization method based on the nuclear norm of the learned features for domain generalization. Intuitively, the proposed regularizer mitigates the impacts of environmental features and encourages learning domain-invariant features. Theoretically, we provide insights into why nuclear norm regularization is more effective compared to ERM and alternative regularization methods. Empirically, we conduct extensive experiments on both synthetic and real datasets. We show nuclear norm regularization achieves strong performance compared to baselines in a wide range of domain generalization tasks. Moreover, our regularizer is broadly applicable with various methods such as ERM and SWAD with consistently improved performance, e.g., 1.7% and 0.9% test accuracy improvements respectively on the DomainBed benchmark. | [
"Domain Generalization",
"Nuclear Norm",
"Deep Learning"
] | https://openreview.net/pdf?id=hJd66ZzXEZ | WRWOaMCLWT | meta_review | 1,699,834,445,911 | hJd66ZzXEZ | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission5/Area_Chair_dkYz"
] | metareview: This paper proposes nuclear norm regularization as a method for domain generalization, and provides empirical and theoretical justifications. The reviewers appreciate the contributions of this paper and unanimously recommend acceptance. One issue pointed out by the reviewers is the sensitivity to hyperparameters, which the authors acknowledge in their response.
recommendation: Accept (Poster)
confidence: 5: The area chair is absolutely certain |
hJd66ZzXEZ | Domain Generalization via Nuclear Norm Regularization | [
"Zhenmei Shi",
"Yifei Ming",
"Ying Fan",
"Frederic Sala",
"Yingyu Liang"
] | The ability to generalize to unseen domains is crucial for machine learning systems deployed in the real world, especially when we only have data from limited training domains. In this paper, we propose a simple and effective regularization method based on the nuclear norm of the learned features for domain generalization. Intuitively, the proposed regularizer mitigates the impacts of environmental features and encourages learning domain-invariant features. Theoretically, we provide insights into why nuclear norm regularization is more effective compared to ERM and alternative regularization methods. Empirically, we conduct extensive experiments on both synthetic and real datasets. We show nuclear norm regularization achieves strong performance compared to baselines in a wide range of domain generalization tasks. Moreover, our regularizer is broadly applicable with various methods such as ERM and SWAD with consistently improved performance, e.g., 1.7% and 0.9% test accuracy improvements respectively on the DomainBed benchmark. | [
"Domain Generalization",
"Nuclear Norm",
"Deep Learning"
] | https://openreview.net/pdf?id=hJd66ZzXEZ | RWPCDBGbeB | official_review | 1,696,707,300,025 | hJd66ZzXEZ | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission5/Reviewer_sUq5"
] | title: Simple approach with thorough evaluations.
review: Summary:
This paper proposes a nuclear norm regularizer to facilitate domain generalization. The proposed approach is evaluated on standard domain generalization benchmarks. Theoretical analysis with a toy example is given to justify the approach.
Advantages:
1. The paper is clearly written and easy to understand. The domain generalization problem is important and relevant.
2. The proposed regularizer is well-motivated with theoretical insights, and the synthetical data experiment offers clear intuition.
3. The method is thoroughly evaluated and compared with existing domain generalization baselines. Although the method itself is not the most performant approach, it shows compatibility with other approaches thanks to its simplicity.
4. The method is ablated carefully, giving the reader more insights into the method's property.
Downsides:
1. While the approach is simple (which is a commendable property, in my opinion), it can incur significant computation burden and instability, placing constraints on the batch size and the representation dimensionality.
2. The performance seems highly sensitive to the weighting factor $\lambda$, which may compromise the practicality of this method. Clearly, there is a subtle balance between norm regularization and classification performance, where the former, if not properly tuned, can throw away valuable information in the model representation.
rating: 6: Marginally above acceptance threshold
confidence: 3: The reviewer is fairly confident that the evaluation is correct |
hJd66ZzXEZ | Domain Generalization via Nuclear Norm Regularization | [
"Zhenmei Shi",
"Yifei Ming",
"Ying Fan",
"Frederic Sala",
"Yingyu Liang"
] | The ability to generalize to unseen domains is crucial for machine learning systems deployed in the real world, especially when we only have data from limited training domains. In this paper, we propose a simple and effective regularization method based on the nuclear norm of the learned features for domain generalization. Intuitively, the proposed regularizer mitigates the impacts of environmental features and encourages learning domain-invariant features. Theoretically, we provide insights into why nuclear norm regularization is more effective compared to ERM and alternative regularization methods. Empirically, we conduct extensive experiments on both synthetic and real datasets. We show nuclear norm regularization achieves strong performance compared to baselines in a wide range of domain generalization tasks. Moreover, our regularizer is broadly applicable with various methods such as ERM and SWAD with consistently improved performance, e.g., 1.7% and 0.9% test accuracy improvements respectively on the DomainBed benchmark. | [
"Domain Generalization",
"Nuclear Norm",
"Deep Learning"
] | https://openreview.net/pdf?id=hJd66ZzXEZ | 5aRyCpNd7O | official_review | 1,696,901,887,918 | hJd66ZzXEZ | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission5/Reviewer_Y5YN"
] | title: nuclear norm regularization
review: This paper studies domain generalization via nuclear norm regularization.
**Quality:**
The paper is technically sound with extensive experiments and some theoretical analysis. The method is simple but effective, and results are strong across diverse datasets.
**Clarity:**
The paper is well-written and easy to follow. The method and experiments are clearly explained.
**Originality:**
Applying nuclear norm regularization in this context is novel and provides a useful regularization technique for domain generalization.
**Significance:**
This technique could have a practical impact given the strong empirical results. The theory also provides new insights.
## Pros
1. Simple and efficient method that consistently improves over baselines.
2. Strong performance across a wide range of benchmark datasets.
3. Theoretical analysis offers insights into benefits of nuclear norm.
4. Easy to implement and combine with other methods.
5. No need for domain labels or changes to model architecture.
## Cons
1. Theoretical assumptions are restrictive; analysis is only for linear models.
2. Gains over existing methods are incremental, not a dramatic breakthrough.
3. Needs some hyperparameter tuning of regularization weight.
4. Sensitivity to hyperparameters could be analyzed more.
5. More analysis to connect theory to deep networks would be beneficial.
rating: 8: Top 50% of accepted papers, clear accept
confidence: 3: The reviewer is fairly confident that the evaluation is correct |
g7rMSiNtmA | Investigating the Catastrophic Forgetting in Multimodal Large Language Model Fine-Tuning | [
"Yuexiang Zhai",
"Shengbang Tong",
"Xiao Li",
"Mu Cai",
"Qing Qu",
"Yong Jae Lee",
"Yi Ma"
] | Following the success of GPT4, there has been a surge in interest in multimodal large language model (MLLM) research. This line of research focuses on developing general-purpose LLMs through fine-tuning pre-trained LLMs and vision models. However, catastrophic forgetting, a notorious phenomenon where the fine-tuned model fails to retain similar performance compared to the pre-trained model, still remains an inherited problem in multimodal LLMs (MLLM). In this paper, we introduce EMT: Evaluating MulTimodality for evaluating the catastrophic forgetting in MLLMs, by treating each MLLM as an image classifier. We first apply EMT to evaluate several open-source fine-tuned MLLMs and we discover that almost all evaluated MLLMs fail to retain the same performance levels as their vision encoders on standard image classification tasks. Moreover, we continue fine-tuning LLaVA, an MLLM and utilize EMT to assess performance throughout the fine-tuning. Interestingly, our results suggest that early-stage fine-tuning on an image dataset improves performance across other image datasets, by enhancing the alignment of text and language features. However, as fine-tuning proceeds, the MLLMs begin to hallucinate, resulting in a significant loss of generalizability, even when the image encoder remains frozen. Our results suggest that MLLMs have yet to demonstrate performance on par with their vision models on standard image classification tasks and the current MLLM fine-tuning procedure still has room for improvement. | [
"Multimodal LLM",
"Supervised Fine-Tuning",
"Catastrophic Forgetting"
] | https://openreview.net/pdf?id=g7rMSiNtmA | yVgk52Qo2v | decision | 1,700,363,269,318 | g7rMSiNtmA | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Program_Chairs"
] | decision: Accept (Oral)
comment: The paper introduces the EMT framework for evaluating catastrophic forgetting in multimodal large language models (MLLMs). It treats MLLMs as image classifiers and shows that early-stage fine-tuning enhances performance across image datasets, but prolonged fine-tuning induces hallucinations, limiting generalizability. Reviewers appreciate the innovative EMT method and the insight into fine-tuning but suggest further exploration of its applicability to diverse multimodal tasks, providing concrete solutions for the identified issues, and a more in-depth analysis of why hallucinations occur. Additionally, addressing reproducibility concerns and experimenting with learning rate strategies during transitions are suggested. Overall, the paper offers valuable contributions to understanding catastrophic forgetting in MLLMs but could benefit from more depth in understanding why and how these hallucinations occur.
The action PC chair for this paper is Atlas Wang, who made the decision after carefully reading the paper as well as the comments by all reviewers and AC. The decision is agreed by all PC chairs.
title: Paper Decision |
g7rMSiNtmA | Investigating the Catastrophic Forgetting in Multimodal Large Language Model Fine-Tuning | [
"Yuexiang Zhai",
"Shengbang Tong",
"Xiao Li",
"Mu Cai",
"Qing Qu",
"Yong Jae Lee",
"Yi Ma"
] | Following the success of GPT4, there has been a surge in interest in multimodal large language model (MLLM) research. This line of research focuses on developing general-purpose LLMs through fine-tuning pre-trained LLMs and vision models. However, catastrophic forgetting, a notorious phenomenon where the fine-tuned model fails to retain similar performance compared to the pre-trained model, still remains an inherited problem in multimodal LLMs (MLLM). In this paper, we introduce EMT: Evaluating MulTimodality for evaluating the catastrophic forgetting in MLLMs, by treating each MLLM as an image classifier. We first apply EMT to evaluate several open-source fine-tuned MLLMs and we discover that almost all evaluated MLLMs fail to retain the same performance levels as their vision encoders on standard image classification tasks. Moreover, we continue fine-tuning LLaVA, an MLLM and utilize EMT to assess performance throughout the fine-tuning. Interestingly, our results suggest that early-stage fine-tuning on an image dataset improves performance across other image datasets, by enhancing the alignment of text and language features. However, as fine-tuning proceeds, the MLLMs begin to hallucinate, resulting in a significant loss of generalizability, even when the image encoder remains frozen. Our results suggest that MLLMs have yet to demonstrate performance on par with their vision models on standard image classification tasks and the current MLLM fine-tuning procedure still has room for improvement. | [
"Multimodal LLM",
"Supervised Fine-Tuning",
"Catastrophic Forgetting"
] | https://openreview.net/pdf?id=g7rMSiNtmA | q4CZzKmcfO | official_review | 1,696,904,983,221 | g7rMSiNtmA | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission41/Reviewer_2R6f"
] | title: Review: Evaluating Catastrophic Forgetting in Multimodal Large Language Models
review: The paper introduces EMT, a novel method for evaluating catastrophic forgetting in multimodal large language models (MLLMs) by treating them as image classifiers, revealing that while early-stage fine-tuning of MLLMs can enhance performance across various image datasets by aligning text and language features, prolonged fine-tuning tends to induce hallucinations in the models, ultimately limiting their generalizability and indicating room for improvement in current MLLM fine-tuning methodologies.
Pros:
1. The EMT method is an innovative approach to assess catastrophic forgetting in MLLMs, providing a different lens by evaluating them as image classifiers
2. The paper provides a interesting and crucial insight of fine-tuning: early-stage fine-tuning appears to enhance performance across various image datasets, while later fine-tuning can make MLLMs hallucinate, diminishing their generalizability even when the image encoder is frozen.
3. The paper is well-written and nicely presented. The experiments are thorough and extensive. This paper opens avenues for subsequent research.
Cons：
1. While EMT offers valuable insights, its applicability and transferability to varied multimodal tasks beyond image classification (segmentation/detection) may need further validation.
2. While the paper expertly identifies and diagnoses issues related to catastrophic forgetting and hallucination in MLLMs, it may lack in offering concrete, actionable solutions or mitigation strategies to address these identified issues.
3. The phenomenon of models beginning to "hallucinate" during fine-tuning is deeply intriguing and could be elaborated further. The paper might benefit from a more in-depth analysis, exploring why and how these hallucinations occur and the intrinsic factors within the MLLMs that contribute to this issue.
rating: 7: Good paper, accept
confidence: 3: The reviewer is fairly confident that the evaluation is correct |
g7rMSiNtmA | Investigating the Catastrophic Forgetting in Multimodal Large Language Model Fine-Tuning | [
"Yuexiang Zhai",
"Shengbang Tong",
"Xiao Li",
"Mu Cai",
"Qing Qu",
"Yong Jae Lee",
"Yi Ma"
] | Following the success of GPT4, there has been a surge in interest in multimodal large language model (MLLM) research. This line of research focuses on developing general-purpose LLMs through fine-tuning pre-trained LLMs and vision models. However, catastrophic forgetting, a notorious phenomenon where the fine-tuned model fails to retain similar performance compared to the pre-trained model, still remains an inherited problem in multimodal LLMs (MLLM). In this paper, we introduce EMT: Evaluating MulTimodality for evaluating the catastrophic forgetting in MLLMs, by treating each MLLM as an image classifier. We first apply EMT to evaluate several open-source fine-tuned MLLMs and we discover that almost all evaluated MLLMs fail to retain the same performance levels as their vision encoders on standard image classification tasks. Moreover, we continue fine-tuning LLaVA, an MLLM and utilize EMT to assess performance throughout the fine-tuning. Interestingly, our results suggest that early-stage fine-tuning on an image dataset improves performance across other image datasets, by enhancing the alignment of text and language features. However, as fine-tuning proceeds, the MLLMs begin to hallucinate, resulting in a significant loss of generalizability, even when the image encoder remains frozen. Our results suggest that MLLMs have yet to demonstrate performance on par with their vision models on standard image classification tasks and the current MLLM fine-tuning procedure still has room for improvement. | [
"Multimodal LLM",
"Supervised Fine-Tuning",
"Catastrophic Forgetting"
] | https://openreview.net/pdf?id=g7rMSiNtmA | gIFqhL8Z3g | official_review | 1,696,713,513,886 | g7rMSiNtmA | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission41/Reviewer_DnDM"
] | title: Investigation of catastrophic forgetting in MLLMs
review: This paper introduces the EMT framework, an approach designed to assess the phenomenon of catastrophic forgetting resulting from the fine-tuning of multimodal large language models (MLLMs). The underlying concept involves treating MLLMs as image classifiers and subjecting them to varying degrees of fine-tuning.
The central finding of this study is the observation that moderate fine-tuning on one dataset confers benefits to datasets not utilized during the fine-tuning process, while excessive fine-tuning leads to catastrophic forgetting. This conclusion aligns seamlessly with logical expectations. When one heavily fine-tunes a model on a specific dataset, the anticipated outcome is an exceptionally high level of accuracy on that specific dataset. There exists no inherent mechanism within the fine-tuning process to incentivize the model to retain optimal performance on non-fine-tuned datasets.
In essence, it begs the question: Why should we anticipate any different outcome?
A few comments and questions for authors:
1. Regarding Section 4.2, titled "Reasons for the Performance Degradation," it appears that this section delves into an investigation and speculation of the factors contributing to failed predictions. However, it lacks evidence establishing a causal link between these identified reasons and the observed decline in model performance. To enhance clarity, I suggest renaming this section to something like "Analyzing Failure Modes of MLLMs."
2. In my humble opinion, fine-tuning models with billions of parameters (e.g., 7B and 13B) on datasets such as CIFAR10 and MNIST may raise concerns about the reliability of the experimental setup.
3. Have the authors explored techniques rooted in gradual unfreezing of layers, such as ULMFiT? These techniques are specifically designed to mitigate the issue of catastrophic forgetting during fine-tuning on downstream datasets. It would be insightful to know whether these methods were considered in the experimentation process.
4. In Section 3.1, would it be advantageous to employ two distinct classification heads for subsets of datasets with non-overlapping classes? The current approach, which adjusts the weights of classes from the pre-trained dataset based on their presence in fine-tuning datasets, may inadvertently encourage the classifier head to prioritize classes currently present in the dataset. This approach lacks a mechanism for preserving features learned for classes from the pre-trained dataset.
5. Could the authors provide clarification on the meaning of "less potent than LLaMa"?
6. The authors mention the use of the OpenAI-API for evaluating results. It is worth noting that this choice may present issues related to the reproducibility of the paper's findings over time, and I would strongly suggest employing an additional (reproducible) method for evaluation purposes.
7. In Section 3.1, did the authors experiment with restarting the learning rate when transitioning from the pre-trained dataset to the fine-tuned dataset? Such a strategy can impact the model's adaptation to new data, and it would be interesting to know if it was explored in the study.
rating: 6
confidence: 2 |
g7rMSiNtmA | Investigating the Catastrophic Forgetting in Multimodal Large Language Model Fine-Tuning | [
"Yuexiang Zhai",
"Shengbang Tong",
"Xiao Li",
"Mu Cai",
"Qing Qu",
"Yong Jae Lee",
"Yi Ma"
] | Following the success of GPT4, there has been a surge in interest in multimodal large language model (MLLM) research. This line of research focuses on developing general-purpose LLMs through fine-tuning pre-trained LLMs and vision models. However, catastrophic forgetting, a notorious phenomenon where the fine-tuned model fails to retain similar performance compared to the pre-trained model, still remains an inherited problem in multimodal LLMs (MLLM). In this paper, we introduce EMT: Evaluating MulTimodality for evaluating the catastrophic forgetting in MLLMs, by treating each MLLM as an image classifier. We first apply EMT to evaluate several open-source fine-tuned MLLMs and we discover that almost all evaluated MLLMs fail to retain the same performance levels as their vision encoders on standard image classification tasks. Moreover, we continue fine-tuning LLaVA, an MLLM and utilize EMT to assess performance throughout the fine-tuning. Interestingly, our results suggest that early-stage fine-tuning on an image dataset improves performance across other image datasets, by enhancing the alignment of text and language features. However, as fine-tuning proceeds, the MLLMs begin to hallucinate, resulting in a significant loss of generalizability, even when the image encoder remains frozen. Our results suggest that MLLMs have yet to demonstrate performance on par with their vision models on standard image classification tasks and the current MLLM fine-tuning procedure still has room for improvement. | [
"Multimodal LLM",
"Supervised Fine-Tuning",
"Catastrophic Forgetting"
] | https://openreview.net/pdf?id=g7rMSiNtmA | VmdZsQ6keB | meta_review | 1,700,406,044,392 | g7rMSiNtmA | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission41/Area_Chair_9vK2"
] | metareview: This paper proposes a novel framework, EMT, to evaluate the catastrophic forgetting phenomenon in multimodal large language models (MLLMs) by treating each of them as an image classifier. Interestingly, they demonstrate that despite early improvements in aligning text and image features, extended fine-tuning results in reduced generalizability and performance issues, highlighting the need for improvement in MLLM development and fine-tuning methods.
Most reviewers agree that EMT is innovative and the observation of models beginning to "hallucinate" during fine-tuning is intriguing. Although some reviewers raise concerns on the limitations of EMT, e.g., only constraints to classification, limited contributions to theoretical analysis or approaches for catastrophic forgetting, rediscovering and evaluating a classical phenomenon in modern multimodal finetuning setting is still widely interesting to CPAL audience and community. Therefore, I recommend the acceptance of this paper.
recommendation: Accept (Poster)
confidence: 4: The area chair is confident but not absolutely certain |
g7rMSiNtmA | Investigating the Catastrophic Forgetting in Multimodal Large Language Model Fine-Tuning | [
"Yuexiang Zhai",
"Shengbang Tong",
"Xiao Li",
"Mu Cai",
"Qing Qu",
"Yong Jae Lee",
"Yi Ma"
] | Following the success of GPT4, there has been a surge in interest in multimodal large language model (MLLM) research. This line of research focuses on developing general-purpose LLMs through fine-tuning pre-trained LLMs and vision models. However, catastrophic forgetting, a notorious phenomenon where the fine-tuned model fails to retain similar performance compared to the pre-trained model, still remains an inherited problem in multimodal LLMs (MLLM). In this paper, we introduce EMT: Evaluating MulTimodality for evaluating the catastrophic forgetting in MLLMs, by treating each MLLM as an image classifier. We first apply EMT to evaluate several open-source fine-tuned MLLMs and we discover that almost all evaluated MLLMs fail to retain the same performance levels as their vision encoders on standard image classification tasks. Moreover, we continue fine-tuning LLaVA, an MLLM and utilize EMT to assess performance throughout the fine-tuning. Interestingly, our results suggest that early-stage fine-tuning on an image dataset improves performance across other image datasets, by enhancing the alignment of text and language features. However, as fine-tuning proceeds, the MLLMs begin to hallucinate, resulting in a significant loss of generalizability, even when the image encoder remains frozen. Our results suggest that MLLMs have yet to demonstrate performance on par with their vision models on standard image classification tasks and the current MLLM fine-tuning procedure still has room for improvement. | [
"Multimodal LLM",
"Supervised Fine-Tuning",
"Catastrophic Forgetting"
] | https://openreview.net/pdf?id=g7rMSiNtmA | VWgb0YKVt7 | official_review | 1,697,403,683,878 | g7rMSiNtmA | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission41/Reviewer_QqLH"
] | title: Review: This paper provides an empirical analysis of catastrophic forgetting in MLLM that is experimentally rich but lacks clear theoretical analysis/support/advancement, which may limit its broader relevance.
review: **Quality:**
The submission demonstrates a methodical approach to understanding catastrophic forgetting in MLLMs through the introduction of the EMT framework. The empirical validation is well-executed with examinations of several open-source fine-tuned MLLMs and a deeper dive into fine-tuning LLaVA.
**Clarity:** The paper is structured in a manner that elucidates the problem, the proposed solution, and the evaluation method which enhances its clarity.
**Originality:**
(a) The introduction of the EMT framework appears to be a novel contribution to the field;
(b) The paper delves into an area that seems less explored in the context of MLLMs, which adds a degree of originality to the work.
**Significance:** The insights garnered from the study, particularly around the effects of fine-tuning, are pertinent and could be instrumental in guiding future work in MLLM fine-tuning.
**Pros:**
- Rich experimental design providing a robust empirical analysis of catastrophic forgetting in MLLMs.
- Introduction of the EMT framework as a novel method to evaluate catastrophic forgetting in MLLMs.
- The paper addresses a relevant and challenging problem in the domain, which could stimulate further research.
**Cons:**
- The work may benefit from a stronger theoretical foundation to support the empirical findings.
- The significance of the contribution may not be very clear or substantial, as noted, and is heavily skewed towards an empirical study rather than novel theoretical or practical contributions.
- A more detailed exploration of the EMT framework's post-processing methods and their implications could provide a fuller understanding of the evaluation process.
rating: 6: Marginally above acceptance threshold
confidence: 4: The reviewer is confident but not absolutely certain that the evaluation is correct |
g2amnrrMP0 | Deep Self-expressive Learning | [
"Chen Zhao",
"Chun-Guang Li",
"Wei He",
"Chong You"
] | Self-expressive model is a method for clustering data drawn from a union of low-dimensional linear subspaces. It gains a lot of popularity due to its: 1) simplicity, based on the observation that each data point can be expressed as a linear combination of the other data points, 2) provable correctness under broad geometric and statistical conditions, and 3) many extensions for handling corrupted, imbalanced, and large-scale real data. This paper extends the self-expressive model to a Deep sELf-expressiVE model (DELVE) for handling the more challenging case that the data lie in a union of nonlinear manifolds. DELVE is constructed from stacking self-expressive layers, each of which maps each data point to a linear combination of the other data points, and can be trained via minimizing self-expressive losses.
With such a design, the operator, architecture, and training of DELVE have the explicit interpretation of producing progressively linearized representations from the input data in nonlinear manifolds. Moreover, by leveraging existing understanding and techniques for self-expressive models, DELVE has a collection of benefits such as design choice by principles, robustness via specialized layers, and efficiency via specialized optimizers. We demonstrate on image datasets that DELVE can effectively perform data clustering, remove data corruptions, and handle large scale data. | [
"Self-Expressive Model; Subspace Clustering; Manifold Clustering"
] | https://openreview.net/pdf?id=g2amnrrMP0 | tdvInNFZI0 | decision | 1,700,361,944,223 | g2amnrrMP0 | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Program_Chairs"
] | decision: Accept (Oral)
comment: The paper introduces DELVE, a self-expressive model for clustering data from nonlinear manifolds. Reviewers generally find the paper well-written and the idea of extending the linear self-expressive model to the nonlinear regime interesting. The paper receives praise for its clear presentation and empirical evaluation. However, some reviewers express concerns about the lack of theoretical guarantees and suggest including stability analysis and standard deviations in the results. Overall, both the reviewers and AC recommend acceptance, recognizing the paper's novelty and high quality.
The action PC chair for this paper is Atlas Wang, who made the decision after carefully reading the paper as well as the comments by all reviewers and AC. The decision is agreed by all PC chairs.
title: Paper Decision |
g2amnrrMP0 | Deep Self-expressive Learning | [
"Chen Zhao",
"Chun-Guang Li",
"Wei He",
"Chong You"
] | Self-expressive model is a method for clustering data drawn from a union of low-dimensional linear subspaces. It gains a lot of popularity due to its: 1) simplicity, based on the observation that each data point can be expressed as a linear combination of the other data points, 2) provable correctness under broad geometric and statistical conditions, and 3) many extensions for handling corrupted, imbalanced, and large-scale real data. This paper extends the self-expressive model to a Deep sELf-expressiVE model (DELVE) for handling the more challenging case that the data lie in a union of nonlinear manifolds. DELVE is constructed from stacking self-expressive layers, each of which maps each data point to a linear combination of the other data points, and can be trained via minimizing self-expressive losses.
With such a design, the operator, architecture, and training of DELVE have the explicit interpretation of producing progressively linearized representations from the input data in nonlinear manifolds. Moreover, by leveraging existing understanding and techniques for self-expressive models, DELVE has a collection of benefits such as design choice by principles, robustness via specialized layers, and efficiency via specialized optimizers. We demonstrate on image datasets that DELVE can effectively perform data clustering, remove data corruptions, and handle large scale data. | [
"Self-Expressive Model; Subspace Clustering; Manifold Clustering"
] | https://openreview.net/pdf?id=g2amnrrMP0 | rnIOfseZ6L | official_review | 1,696,752,071,749 | g2amnrrMP0 | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission42/Reviewer_LuGX"
] | title: Review
review: **Summary**:
The paper proposes a self-expressive model for clustering data drawn from a union of nonlinear manifolds, extending methods working in the linear regime.
**Pros**:
- The paper is, in general, easy to follow and clearly written.
- The proposed method is a natural extension of a similar idea introduced for linear manifolds, and, within the provided in the paper empirical evaluation, is effective.
**Cons**:
- The method seems to be grounded in intuition from linear case. In consequence, I am not sure whether it is theoretically guaranteed to work
- The provided results could be complemented by a comment on the stability of the methods, as well as standard deviations.
- The text ends abruptly
(See the “Details” for more comments/reasoning behind the above pros and cons)
**Details**:
The paper is of good quality and clarity. I also consider the idea of extending the linear approach of self-expressive models to a nonlinear case by progressively linearizing the representations interesting. However, apart from the provided intuition given by the authors on page 2 I do not understand (or see) whether there any (even simplified) theoretical guarantees guiding this approach (not that I necessarily need to see ones, but would appreciate a comment on their existence in the text). I appreciate the provided evaluation with other approaches. However,I would like to have the blank spaces in Table 2 clarified (i.e. why some methods were not evaluated on some data). Additionally, I would also like to see a comment on the stability of the methods, preferably by adding information about deviation from the results reported in the Tables in the text. Finally, In general, I believe the paper is easy to follow and clearly written, however, it ends abruptly, without any remarks on the closing conclusions. I would like to see those issues addressed. Beyond the points above I do not see any major flaws in the paper.
rating: 7: Good paper, accept
confidence: 2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper |
g2amnrrMP0 | Deep Self-expressive Learning | [
"Chen Zhao",
"Chun-Guang Li",
"Wei He",
"Chong You"
] | Self-expressive model is a method for clustering data drawn from a union of low-dimensional linear subspaces. It gains a lot of popularity due to its: 1) simplicity, based on the observation that each data point can be expressed as a linear combination of the other data points, 2) provable correctness under broad geometric and statistical conditions, and 3) many extensions for handling corrupted, imbalanced, and large-scale real data. This paper extends the self-expressive model to a Deep sELf-expressiVE model (DELVE) for handling the more challenging case that the data lie in a union of nonlinear manifolds. DELVE is constructed from stacking self-expressive layers, each of which maps each data point to a linear combination of the other data points, and can be trained via minimizing self-expressive losses.
With such a design, the operator, architecture, and training of DELVE have the explicit interpretation of producing progressively linearized representations from the input data in nonlinear manifolds. Moreover, by leveraging existing understanding and techniques for self-expressive models, DELVE has a collection of benefits such as design choice by principles, robustness via specialized layers, and efficiency via specialized optimizers. We demonstrate on image datasets that DELVE can effectively perform data clustering, remove data corruptions, and handle large scale data. | [
"Self-Expressive Model; Subspace Clustering; Manifold Clustering"
] | https://openreview.net/pdf?id=g2amnrrMP0 | pRKmprKiKJ | official_review | 1,696,549,784,936 | g2amnrrMP0 | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission42/Reviewer_gKan"
] | title: Review of Submission 42: Self-expressive Learning
review: ## Overview:
In this paper, the authors extend self-expressive learning, which can be used to identify separate *linear*, low-dimensional manifolds that data lies in the union of, to the setting where the manifolds are non-linear. This captures a much wider class of data-sets, and through numerical experiments the authors show how their deep self-expressive approach can cluster with state-of-the-art accuracy on real world data like faces (EYaleB) and handwritten symbols (MNIST). Additionally, they show how their approach can handle corruption and can be further modified for other possible settings.
## Strengths:
1. This paper presents a novel approach to self-expressive learning that yields strong numerical results.
2. This paper is well aligned to the overarching goals of CPAL.
3. This paper is well written and easy to follow. I was unfamiliar with self-expressive learning, but the authors enabled me to easily follow along.
4. The illustrations of the approach (Figs. 1 and 2) were very helpful in understanding the method and made it easy to follow along.
5. The authors discuss several ways in which to use in their approach in other settings, such as by modifying the regularizer and penalty term. I believe this will make it easy for others to implement, in an *intelligent* manner.
6. The authors present a comparison of their approach to self-attention and graph convolution networks, which I found to be interesting, informative, and topical.
## Weaknesses:
I found no major weaknesses of the paper.
However, I do think there are a few points that, if addressed. would make this paper stronger:
1. In Fig. 2, DELVE is compared to an autoencoder. While the representations learned by DELVE look much nicer (and support the authors' claims), the comparison to the autoencoder is complicated by the fact that only 4 layers of the autoencoder are shown, as compared to 40 DELVE layers. I assume the reason there aren't more layers of the autoencoder is because it is difficult to train one with 40 layers? Commenting on this rationale, as well as showing perhaps the DELVE representations after 2 or 4 layers (so as to at least match the autoencoder depth once in the figure), would be helpful in making a more clear comparison.
2. It would be helpful for the reader to know more about the data sets used in Sec. 4 (what is COIL100?), as well as the other methods that DELVE is compared to (what is NMCE, and is it known that ``The better performance of NMCE relies critically on its use of data augmentation'' [lines 241-242]? If so, is there a citation that could be referenced?). It would additionally be helpful to reference, in the main text, that there are Tables and extra details presented in the Appendices, so that the reader can know that more details of the experiments are reported elsewhere.
3. The paper ends rather abruptly, as there is no conclusion. I think it is reasonable to move some of Table 2 to the supplement and then have a paragraph or two re-iterating why this work is important and what future directions the authors foresee.
## Questions:
1. The choice of penalty term (Eq. 5) appears to be motivated by the elastic net regularizer (Eq. 4), but is never explicitly stated. Is that the correct motivation, or is there a separate reason for choosing that form?
## Summary:
This is a great paper that introduces a novel idea, that fits well within the CPAL CfP and yields strong experimental results. I believe would be a strong addition to the first annual CPAL.
rating: 8: Top 50% of accepted papers, clear accept
confidence: 3: The reviewer is fairly confident that the evaluation is correct |
g2amnrrMP0 | Deep Self-expressive Learning | [
"Chen Zhao",
"Chun-Guang Li",
"Wei He",
"Chong You"
] | Self-expressive model is a method for clustering data drawn from a union of low-dimensional linear subspaces. It gains a lot of popularity due to its: 1) simplicity, based on the observation that each data point can be expressed as a linear combination of the other data points, 2) provable correctness under broad geometric and statistical conditions, and 3) many extensions for handling corrupted, imbalanced, and large-scale real data. This paper extends the self-expressive model to a Deep sELf-expressiVE model (DELVE) for handling the more challenging case that the data lie in a union of nonlinear manifolds. DELVE is constructed from stacking self-expressive layers, each of which maps each data point to a linear combination of the other data points, and can be trained via minimizing self-expressive losses.
With such a design, the operator, architecture, and training of DELVE have the explicit interpretation of producing progressively linearized representations from the input data in nonlinear manifolds. Moreover, by leveraging existing understanding and techniques for self-expressive models, DELVE has a collection of benefits such as design choice by principles, robustness via specialized layers, and efficiency via specialized optimizers. We demonstrate on image datasets that DELVE can effectively perform data clustering, remove data corruptions, and handle large scale data. | [
"Self-Expressive Model; Subspace Clustering; Manifold Clustering"
] | https://openreview.net/pdf?id=g2amnrrMP0 | hqO9K2uv1H | meta_review | 1,699,837,960,773 | g2amnrrMP0 | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission42/Area_Chair_E8wP"
] | metareview: This paper proposes a self-expressive model to cluster data drawn from a union of nonlinear manifold. It is a novel extension of the self-expressive model from the linear regime to the nonlinear regime. All reviewers found the proposed approach interesting and novel, and the proposed DELVE was supported by implementation details and experimental validation. I agree with the reviewers that this paper is of high quality, and recommend an acceptance.
recommendation: Accept (Poster)
confidence: 4: The area chair is confident but not absolutely certain |
g2amnrrMP0 | Deep Self-expressive Learning | [
"Chen Zhao",
"Chun-Guang Li",
"Wei He",
"Chong You"
] | Self-expressive model is a method for clustering data drawn from a union of low-dimensional linear subspaces. It gains a lot of popularity due to its: 1) simplicity, based on the observation that each data point can be expressed as a linear combination of the other data points, 2) provable correctness under broad geometric and statistical conditions, and 3) many extensions for handling corrupted, imbalanced, and large-scale real data. This paper extends the self-expressive model to a Deep sELf-expressiVE model (DELVE) for handling the more challenging case that the data lie in a union of nonlinear manifolds. DELVE is constructed from stacking self-expressive layers, each of which maps each data point to a linear combination of the other data points, and can be trained via minimizing self-expressive losses.
With such a design, the operator, architecture, and training of DELVE have the explicit interpretation of producing progressively linearized representations from the input data in nonlinear manifolds. Moreover, by leveraging existing understanding and techniques for self-expressive models, DELVE has a collection of benefits such as design choice by principles, robustness via specialized layers, and efficiency via specialized optimizers. We demonstrate on image datasets that DELVE can effectively perform data clustering, remove data corruptions, and handle large scale data. | [
"Self-Expressive Model; Subspace Clustering; Manifold Clustering"
] | https://openreview.net/pdf?id=g2amnrrMP0 | bDpoQQQS4M | official_review | 1,697,402,031,636 | g2amnrrMP0 | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission42/Reviewer_yobb"
] | title: Progressive linearization and clustering of data drawn from a union of manifolds
review: The goal of this work is to obtain linearized representations and cluster data that is drawn from a union of low dimensional manifolds. Existing methods either have difficulties with less/noisy samples or use deep models that lack interpretability.
To achieve this task using interpretable deep models, the authors propose an approach called DELVE that takes inspiration from linear subspace clustering methods. The key idea used by DELVE is that a data point on a non-linear manifold can be approximated using a linear combination of points in its local neighborhood on that same manifold. This is used to build a sequence of ‘self-expressive layers’ that progressively linearize the input manifold. The parameters of each layer are sequentially optimized in an unsupervised manner by using a ‘self-expressive loss’.
**Strengths:**
1. DELVE achieves good clustering performance, comparable to existing deep models. The model can also scale well to large datasets since it does not necessarily require backpropagation through all the layers. Additionally, from Fig. 2, the obtained representations visually appear meaningful.
2. I like the white-box characteristic of DELVE. The role played by each layer is interpretable and one can play with different hyper-parameters and observe performance changes from the model that are mathematically explainable.
3. The paper is generally well written.
I have a few questions/concerns regarding DELVE that I would like the authors to clarify.
1. A key premise for DELVE to succeed is the ability to linearly approximate a data point using other points in its local neighborhood on the manifold. How would DELVE perform in comparison to an autoencoder if the ambient dimension/image size increases (or available data points from the manifold are farther apart from each other)? How do the obtained representations in such case look like? Are they still meaningful?
One of the limitations of conventional approaches like locally linear embeddings is their difficulty to deal with less samples. I wonder if DELVE would have a similar performance decline (if any) or its model depth could somehow help with that.
2. Once an autoencoder is trained, it can perform a quick forward pass to yield representations for a new data point. On the other hand, it would be quite inefficient if DELVE had to run the entire optimization program every time one needed representation for a new data point. How quickly can DELVE do this in comparison to a conventional autoencoder? Is there an efficient method that DELVE can utilize?
rating: 7
confidence: 2 |
cyMsUO5J7U | Sparse Activations with Correlated Weights in Cortex-Inspired Neural Networks | [
"Chanwoo Chun",
"Daniel Lee"
] | Although sparse activations are commonly seen in cortical brain circuits, the computational benefits of sparse activations are not well understood for machine learning. Recent neural network Gaussian Process models have incorporated sparsity in infinitely-wide neural network architectures, but these models result in Gram matrices that approach the identity matrix with increasing sparsity. This collapse of input pattern similarities in the network representation is due to the use of independent weight vectors in the models. In this work, we show how weak correlations in the weights can counter this effect. Correlations in the synaptic weights are introduced using a convolutional model, similar to the neural structure of lateral connections in the cortex. We show how to theoretically compute the properties of infinitely-wide networks with sparse, correlated weights and with rectified linear outputs. In particular, we demonstrate how the generalization performance of these sparse networks improves by introducing these correlations. We also show how to compute the optimal degree of correlations that result in the best-performing deep networks. | [
"Correlated weights",
"Biological neural network",
"Cortex",
"Neural network gaussian process",
"Sparse neural network",
"Bayesian neural network",
"Generalization theory",
"Kernel ridge regression",
"Deep neural network",
"Random neural network"
] | https://openreview.net/pdf?id=cyMsUO5J7U | v4zcPSPnAd | official_review | 1,696,633,852,187 | cyMsUO5J7U | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission22/Reviewer_FTVf"
] | title: Counteracting sparsity with weak correlations
review: ### Summary :
This work suggests weakly correlated weights in the neural network Gaussian Process model enable learning useful sparse representations. Empirical evaluation demonstrates that weak correlations (via an intermediate convolutional layer) induce low-dimensional kernel Gram matrix in the sparse regime. Theoretical justification is also provided for the same observation which in turn predicts good generalization behaviour. Informally sparsity dissimilates the neural representations across layers while correlations counter this effect. For any sparsity level f, the authors derive an optimal correlation level that balances the tradeoff in generalization.
rating: 7: Good paper, accept
confidence: 1: The reviewer's evaluation is an educated guess |
cyMsUO5J7U | Sparse Activations with Correlated Weights in Cortex-Inspired Neural Networks | [
"Chanwoo Chun",
"Daniel Lee"
] | Although sparse activations are commonly seen in cortical brain circuits, the computational benefits of sparse activations are not well understood for machine learning. Recent neural network Gaussian Process models have incorporated sparsity in infinitely-wide neural network architectures, but these models result in Gram matrices that approach the identity matrix with increasing sparsity. This collapse of input pattern similarities in the network representation is due to the use of independent weight vectors in the models. In this work, we show how weak correlations in the weights can counter this effect. Correlations in the synaptic weights are introduced using a convolutional model, similar to the neural structure of lateral connections in the cortex. We show how to theoretically compute the properties of infinitely-wide networks with sparse, correlated weights and with rectified linear outputs. In particular, we demonstrate how the generalization performance of these sparse networks improves by introducing these correlations. We also show how to compute the optimal degree of correlations that result in the best-performing deep networks. | [
"Correlated weights",
"Biological neural network",
"Cortex",
"Neural network gaussian process",
"Sparse neural network",
"Bayesian neural network",
"Generalization theory",
"Kernel ridge regression",
"Deep neural network",
"Random neural network"
] | https://openreview.net/pdf?id=cyMsUO5J7U | L7gpOQOMZB | official_review | 1,696,767,619,536 | cyMsUO5J7U | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission22/Reviewer_sLsf"
] | title: An interesting work discovering the correlation between weight correlation and activation sparsity within infinitely-wide networks.
review: This paper focuses on the benefits of sparse activations in infinitely-wide networks and studies how weak correlations in the weights can improve the generalization performance. The weights have been typically assumed to be independent so understanding the effect of correlated weights is currently limited. To this end, this paper first proposed a variant of NNGP for correlated weights, showing that the inducing of correlated wights is capable of improving the generalization performance in the sparse regime. The corresponding theoretical explanation is further provided by extending the recent advances in generalization theory. Eventually, the optimal weight correlation give a target sparsity has been introduced, improving the practical usage of this paper.
Overall, this paper is presented with high quality and clarity. I enjoy much when reading the introduction and review of sparse NNGP.
The novelty of this paper mainly lies on (1) the formulation on extending of NNGP for correlated weights; (2) the empirical evidence for the benefits of weights correlation to feature sparsity, and to generalization performance; (3) Theoretical proof.
My major concern is why this paper mainly focus on random weighted sparse networks? Can the findings be generalized to trained sparse networks, the most closely one is "The lazy neuron phenomenon: On emergence
318 of activation sparsity in transformers"; and sparse training regime i.e., SET (https://www.nature.com/articles/s41467-018-04316-3), SNIP (https://arxiv.org/abs/1810.02340), ITOP (https://arxiv.org/abs/2102.02887).
I also list several minor cons of this manuscript here:
(1) The authors mention that "Therefore, sparse random models do not perform well with deep architectures [12]." However, IMHO, there are empirical papers showing that when models get deeper and wider, sparse random models actually perform better, i.e., random pruning https://arxiv.org/abs/2202.02643, and random sparse GNNs: https://arxiv.org/abs/2211.15335. While the randomness in this manuscript is slightly different from the papers I mentioned, it is necessary to at least discuss and explain this two seemingly counter-arguments.
(2) I encourage the authors provide a contribution summary in the early of the paper to improve the readiness of this paper.
(3) When this paper mentions in the intro "Currently, there is no theoretical explanation of how deep network models such as the Transformers benefit from sparsity" as their motivation, I expect to see any analysis of Transformers in the main paper. However, it is not presented. I encourage the authors to say something about Transformers or simply remove this statement.
rating: 6: Marginally above acceptance threshold
confidence: 3: The reviewer is fairly confident that the evaluation is correct |
cyMsUO5J7U | Sparse Activations with Correlated Weights in Cortex-Inspired Neural Networks | [
"Chanwoo Chun",
"Daniel Lee"
] | Although sparse activations are commonly seen in cortical brain circuits, the computational benefits of sparse activations are not well understood for machine learning. Recent neural network Gaussian Process models have incorporated sparsity in infinitely-wide neural network architectures, but these models result in Gram matrices that approach the identity matrix with increasing sparsity. This collapse of input pattern similarities in the network representation is due to the use of independent weight vectors in the models. In this work, we show how weak correlations in the weights can counter this effect. Correlations in the synaptic weights are introduced using a convolutional model, similar to the neural structure of lateral connections in the cortex. We show how to theoretically compute the properties of infinitely-wide networks with sparse, correlated weights and with rectified linear outputs. In particular, we demonstrate how the generalization performance of these sparse networks improves by introducing these correlations. We also show how to compute the optimal degree of correlations that result in the best-performing deep networks. | [
"Correlated weights",
"Biological neural network",
"Cortex",
"Neural network gaussian process",
"Sparse neural network",
"Bayesian neural network",
"Generalization theory",
"Kernel ridge regression",
"Deep neural network",
"Random neural network"
] | https://openreview.net/pdf?id=cyMsUO5J7U | Kv7y65GtnB | official_review | 1,696,704,074,561 | cyMsUO5J7U | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission22/Reviewer_qzTS"
] | title: Review of Sparse Activations with Correlated Weights in Cortex-Inspired Neural Networks
review: ### Summary of contributions:
This work demonstrates, both theoretically and empirically, how adding weak correlations between weights in an infinitely-wide neural network can counter the dissimilating effect of sparse activations which in turn improves generalization performance on three small scale datasets.
### Strengths:
1. This is a timely and important topic as sparse activations are observed in many neural network architectures and in particular the now ubiquitous transformer architecture.
2. The paper is well motivated based on prior art and biologically similar mechanisms within cortical circuits.
3. The study of using correlated weights in a NNGP setting appears to be novel.
4. The formula presented for calculating the optimal weight correlation is of particular interest. Future work exploring applications of correlated weights may be able to realize modest generalization performance improvements by implementing correlated weights.
5. Empirical results generally match very closely with results predicted by theoretical results, particularly for experiments with smaller training sets.
### Weaknesses:
1. Experiments with m=20 do appear to consistently yield the best generalization performance, especially at moderate to high sparsities with small training datasets (P<=512). However, as P is increased, the gap between m=0 and m=20 appears to narrow, particularly for CIFAR and MNIST. I also note that the gap between theoretical predictions and empirical results in Figure S4 appears to increase proportionally with P. A more detailed discussion on the effect of increasing training set size would improve the paper.
3. While typical of the NNGP literature, the datasets used for the analysis are small and the training set sizes used are very small for typical training of neural networks. It is unclear if the weight correlation benefits will remain if larger, real-world datasets are used. Did the authors conduct any experiments using training sets larger than 4096? If not, what was the reason for excluding larger numbers of training samples? I remain skeptical if benefits observed with weight correlation will continue to be present as the training set size is scaled up.
4. The empirical study is limited to the NNGP setting. Experiments with finite-width networks in addition to the kernel method experiments would help establish whether the proposed weight correlation methodology is broadly applicable to real-world neural networks.
5. For MNIST and FashionMNIST, the benefits of weight correlation appear to be most prominent in the highly sparse regime. However, the best absolute generalization performance is typically found at modest sparsities where the benefits of using correlated weights are less compelling.
6. Only a single type of network architecture was considered for the empirical study. Additional experiments with deeper models would improve confidence in the results.
### Clarifications:
1. "More recently, the presence of sparse activation has been observed in high-performance neural networks such as AlexNet, **ImageNet**, LeNet, and various models of Transformers, even without explicit regularization for sparsity."ImageNet" typically refers to the ILSVR challenge datasets and not a neural network architecture. Please confirm intended network.
2. "f" in equation 5 refers to the "sparsity level". However, I believe the actual intent of this variable is to represent the "fixed fraction of neurons with non-zero activations" as per [1]. I would expect sparsity level to be defined as (1-f) if this is correct. Please clarify.
3. Figure 3 caption states that "Red triangle marks f = 0.2, m = 20 as the best-performing model.". This seems like a reach to me as essentially all correlation levels have the same performance at that sparsity. What is the absolute difference between m=0 and m=20 at the red triangle? If it's as small as it appears in the plot, I recommend removing this statement as it seems disingenuous.
4. On line 192, P is defined as "the size of the training set". This is clearly established in plots throughout the paper, but seems to be placed somewhat awkwardly with respect to equation 13 since P does not appear in that equation.
### Suggested changes:
1. Adding hyperlinks to the in-text citations that link to the bibliography would be helpful.
2. Move definition of P to a more appropriate location, preferably near Figure 2 when it is first used in caption and figure plot titles.
3. In Figure 4b), m=150 series is hard to read. Consider use of a different color.
### Broader impact concerns:
None.
### Citations
[1] C. Chun and D. D. Lee, “Sparsity-depth Tradeoff in Infinitely Wide Deep Neural Networks.” arXiv, May 17, 2023. doi: 10.48550/arXiv.2305.10550.
rating: 7: Good paper, accept
confidence: 3: The reviewer is fairly confident that the evaluation is correct |
cyMsUO5J7U | Sparse Activations with Correlated Weights in Cortex-Inspired Neural Networks | [
"Chanwoo Chun",
"Daniel Lee"
] | Although sparse activations are commonly seen in cortical brain circuits, the computational benefits of sparse activations are not well understood for machine learning. Recent neural network Gaussian Process models have incorporated sparsity in infinitely-wide neural network architectures, but these models result in Gram matrices that approach the identity matrix with increasing sparsity. This collapse of input pattern similarities in the network representation is due to the use of independent weight vectors in the models. In this work, we show how weak correlations in the weights can counter this effect. Correlations in the synaptic weights are introduced using a convolutional model, similar to the neural structure of lateral connections in the cortex. We show how to theoretically compute the properties of infinitely-wide networks with sparse, correlated weights and with rectified linear outputs. In particular, we demonstrate how the generalization performance of these sparse networks improves by introducing these correlations. We also show how to compute the optimal degree of correlations that result in the best-performing deep networks. | [
"Correlated weights",
"Biological neural network",
"Cortex",
"Neural network gaussian process",
"Sparse neural network",
"Bayesian neural network",
"Generalization theory",
"Kernel ridge regression",
"Deep neural network",
"Random neural network"
] | https://openreview.net/pdf?id=cyMsUO5J7U | Ej6FUS0Jt0 | meta_review | 1,699,563,998,364 | cyMsUO5J7U | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission22/Area_Chair_C2co"
] | metareview: The paper investigates the role of sparsity in infinitely-wide Neural Network Gaussian Process (NNGP) models, particularly focusing on the benefits of correlated weights on generalization performance. Strengths highlighted by reviewers include the theoretical advancements in NNGP by looking at correlated weights rather than independent weights, and the proposed optimal weight correlation formulation for specific sparsity targets. Some reviewers expressed concerns regarding the limited scope of experiments, primarily relying on small datasets and training set sizes, and questioned whether the authors insights for random sparse networks carry over to trained sparse networks. However, these issues were addressed in the author rebuttal to the apparent satisfactions of the reviewers. Overall, a strong theoretical work that should have broad appeal to researchers interested in understanding the role sparsity in both biological and artificial neural networks.
recommendation: Accept (Oral)
confidence: 4: The area chair is confident but not absolutely certain |
cyMsUO5J7U | Sparse Activations with Correlated Weights in Cortex-Inspired Neural Networks | [
"Chanwoo Chun",
"Daniel Lee"
] | Although sparse activations are commonly seen in cortical brain circuits, the computational benefits of sparse activations are not well understood for machine learning. Recent neural network Gaussian Process models have incorporated sparsity in infinitely-wide neural network architectures, but these models result in Gram matrices that approach the identity matrix with increasing sparsity. This collapse of input pattern similarities in the network representation is due to the use of independent weight vectors in the models. In this work, we show how weak correlations in the weights can counter this effect. Correlations in the synaptic weights are introduced using a convolutional model, similar to the neural structure of lateral connections in the cortex. We show how to theoretically compute the properties of infinitely-wide networks with sparse, correlated weights and with rectified linear outputs. In particular, we demonstrate how the generalization performance of these sparse networks improves by introducing these correlations. We also show how to compute the optimal degree of correlations that result in the best-performing deep networks. | [
"Correlated weights",
"Biological neural network",
"Cortex",
"Neural network gaussian process",
"Sparse neural network",
"Bayesian neural network",
"Generalization theory",
"Kernel ridge regression",
"Deep neural network",
"Random neural network"
] | https://openreview.net/pdf?id=cyMsUO5J7U | DQxOJatDHX | decision | 1,700,362,935,204 | cyMsUO5J7U | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Program_Chairs"
] | decision: Accept (Oral)
comment: This paper investigates the impact of sparse activations and correlated weights in infinitely-wide Neural Network Gaussian Process (NNGP) models, aiming to improve generalization performance. Reviewers appreciate the theoretical advancements in NNGP related to correlated weights and the formulation of optimal weight correlation for specific sparsity targets. However, some concerns include the limited experimental scope, reliance on small datasets and training set sizes, and questions about the applicability of findings to trained sparse networks. Nevertheless, these concerns were addressed in the author rebuttal. Overall, the paper presents a strong theoretical contribution relevant to researchers interested in sparsity in neural networks, both biological and artificial.
The action PC chair for this paper is Atlas Wang, who made the decision after carefully reading the paper as well as the comments by all reviewers and AC. The decision is agreed by all PC chairs.
title: Paper Decision |
Z0Fk5MyxkY | Piecewise-Linear Manifolds for Deep Metric Learning | [
"Shubhang Bhatnagar",
"Narendra Ahuja"
] | Unsupervised deep metric learning (UDML) focuses on learning a semantic representation space using only unlabeled data. This challenging problem requires accurately estimating the similarity between data points, which is used to supervise a deep network. For this purpose, we propose to model the high-dimensional data manifold using a piecewise-linear approximation, with each low-dimensional linear piece approximating the data manifold in a small neighborhood of a point. These neighborhoods are used to estimate similarity between data points. We empirically show that this similarity estimate correlates better with the ground truth than the similarity estimates of current state-of-the-art techniques. We also show that proxies, commonly used in supervised metric learning, can be used to model the piecewise-linear manifold in an unsupervised setting, helping improve performance. Our method outperforms existing unsupervised metric learning approaches on standard zero-shot image retrieval benchmarks. | [
"Deep metric learning",
"Unsupervised representation learning",
"Image retrieval"
] | https://openreview.net/pdf?id=Z0Fk5MyxkY | rcGJHrEis5 | official_review | 1,696,433,782,214 | Z0Fk5MyxkY | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission37/Reviewer_vgyP"
] | title: Good results, but convoluted method
review: This paper proposed a new algorithm for deep metric learning based on estimating the piecewise-linear local manifold. Nearest-neighbor based sampling and proxy vectors are used to mitigate the issue that a random batch doesn’t contain enough samples to form a local manifold. Compared to other metric learning methods, the proposed method achieves significantly better performance.
Overall, this paper is clear and well written. The method and experimental results are clearly presented. The proposed method is motivated in terms of encouraging sample points to come closer to the local manifold formed by neighboring points, but the resulting loss function seems much more complicated than it needs to be. First, there’s very little motivation provided for the particular form of similarity metric s proposed in Section 3.3, besides transforming the projected distance and orthogonal distance to the required numerical range. There’s also no motivation as of why ɑ and ꞵ needs to be multiplied together in equation (2).
It is also interesting that in the actual experiment, the authors mixed this distance term with other simpler terms like L2 distance and cosine similarity again. This happens in equation (6) (7) and (8). A δ parameter is included to adjust the importance of similarity s and other terms. However, in Figure 3, when δ decreases, it only has a very minor effect on the accuracy result, indicating that the other, simpler terms are doing most of the heavy lifting.
The proposed method also contains many extra implementation details that other metric learning methods do not have, for example, using nearest neighbor sampling to form a batch. In the self-supervised learning literature, it is well known that this alone can provide significant performance improvement[1]. Therefore, it will be necessary to study what design choice contributed to the improved performance and by how much.
Although it is nice that the authors achieved significant improvements in their benchmark results, it is at the expense of many more hyper-parameters involved, as well as a more convoluted and ill-motivated design. It would be more interesting to see experiments and analysis about how this improvement is achieved and what principle it reveals.
[1] Debidatta, Dwibedi. Et al. With a Little Help from My Friends: Nearest-Neighbor Contrastive Learning of Visual Representations. arXiv:2104.14548v1
rating: 5
confidence: 4 |
Z0Fk5MyxkY | Piecewise-Linear Manifolds for Deep Metric Learning | [
"Shubhang Bhatnagar",
"Narendra Ahuja"
] | Unsupervised deep metric learning (UDML) focuses on learning a semantic representation space using only unlabeled data. This challenging problem requires accurately estimating the similarity between data points, which is used to supervise a deep network. For this purpose, we propose to model the high-dimensional data manifold using a piecewise-linear approximation, with each low-dimensional linear piece approximating the data manifold in a small neighborhood of a point. These neighborhoods are used to estimate similarity between data points. We empirically show that this similarity estimate correlates better with the ground truth than the similarity estimates of current state-of-the-art techniques. We also show that proxies, commonly used in supervised metric learning, can be used to model the piecewise-linear manifold in an unsupervised setting, helping improve performance. Our method outperforms existing unsupervised metric learning approaches on standard zero-shot image retrieval benchmarks. | [
"Deep metric learning",
"Unsupervised representation learning",
"Image retrieval"
] | https://openreview.net/pdf?id=Z0Fk5MyxkY | hlg2iNpGbD | decision | 1,700,432,034,684 | Z0Fk5MyxkY | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Program_Chairs"
] | decision: Accept (Oral)
comment: After a careful review of the paper titled "Learning Semantically Meaningful Representations in Unsupervised Deep Metric Learning," the reviewers found the research to be interesting and noted significant improvements. The paper addresses the important problem of learning semantically meaningful representations from unlabeled data, specifically focusing on Unsupervised Deep Metric Learning (UDML) and the challenges posed by noisy pseudo-labels generated through clustering techniques.
The proposed approach of modeling data as piecewise linear manifolds is seen as a novel and promising alternative to traditional clustering methods. It offers advantages in terms of label homogeneity and the estimation of less noisy, continuous-valued similarity. The inclusion of proxies, a technique commonly used in supervised metric learning, is also deemed valuable in an unsupervised setting, leading to improved performance and faster convergence.
Given the positive feedback from the reviewers, We decided to accept the paper. However, there are some remaining technical concerns that the authors should address in their final manuscript. It is crucial for the authors to carefully consider and address these concerns to ensure the quality and rigor of the work.
The action PC chair for this paper is Qing Qu, who made the decision after carefully reading the paper as well as the comments by all reviewers and AC. The decision is agreed upon by all PC chairs.
title: Paper Decision |
Z0Fk5MyxkY | Piecewise-Linear Manifolds for Deep Metric Learning | [
"Shubhang Bhatnagar",
"Narendra Ahuja"
] | Unsupervised deep metric learning (UDML) focuses on learning a semantic representation space using only unlabeled data. This challenging problem requires accurately estimating the similarity between data points, which is used to supervise a deep network. For this purpose, we propose to model the high-dimensional data manifold using a piecewise-linear approximation, with each low-dimensional linear piece approximating the data manifold in a small neighborhood of a point. These neighborhoods are used to estimate similarity between data points. We empirically show that this similarity estimate correlates better with the ground truth than the similarity estimates of current state-of-the-art techniques. We also show that proxies, commonly used in supervised metric learning, can be used to model the piecewise-linear manifold in an unsupervised setting, helping improve performance. Our method outperforms existing unsupervised metric learning approaches on standard zero-shot image retrieval benchmarks. | [
"Deep metric learning",
"Unsupervised representation learning",
"Image retrieval"
] | https://openreview.net/pdf?id=Z0Fk5MyxkY | gUOJCgi6q4 | official_review | 1,696,346,181,866 | Z0Fk5MyxkY | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission37/Reviewer_3LmT"
] | title: Accept with several questions/suggestions
review: The authors present a novel method for unsupervised deep metric learning by representing data with piecewise linear manifolds. This simple representation allows neighborhoods to be linearly approximated.
Pros:
+ Provides good intuition for their approach
+ Clear motivation
+ Validated their approach against numerous other methods
Cons:
- Method description is fairly long and involves many variables, so an algorithm/pseudocode block would enhance the clarity.
- Limits of their work not explicitly stated
- Choice of GoogLeNet, a fairly old model, over other models is not explained
Questions:
1) The loss function is a sum of three components (point, proxy, and neighborhood). Is there a reason to weight these the same i.e. would performance improve by adding the flexibility to weight them differently?
2) What is the reasoning behind choosing GoogLeNet for your experiments? How does the method perform with different architectures and larger models?
rating: 8
confidence: 3 |
Z0Fk5MyxkY | Piecewise-Linear Manifolds for Deep Metric Learning | [
"Shubhang Bhatnagar",
"Narendra Ahuja"
] | Unsupervised deep metric learning (UDML) focuses on learning a semantic representation space using only unlabeled data. This challenging problem requires accurately estimating the similarity between data points, which is used to supervise a deep network. For this purpose, we propose to model the high-dimensional data manifold using a piecewise-linear approximation, with each low-dimensional linear piece approximating the data manifold in a small neighborhood of a point. These neighborhoods are used to estimate similarity between data points. We empirically show that this similarity estimate correlates better with the ground truth than the similarity estimates of current state-of-the-art techniques. We also show that proxies, commonly used in supervised metric learning, can be used to model the piecewise-linear manifold in an unsupervised setting, helping improve performance. Our method outperforms existing unsupervised metric learning approaches on standard zero-shot image retrieval benchmarks. | [
"Deep metric learning",
"Unsupervised representation learning",
"Image retrieval"
] | https://openreview.net/pdf?id=Z0Fk5MyxkY | eIuiKbheYl | official_review | 1,696,713,112,810 | Z0Fk5MyxkY | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission37/Reviewer_cHmX"
] | title: Good result, more qualitative and exploratory analysis would be helpful.
review: Summary: The paper is dealing with unsupervised deep metric learning, which aims to learn a metric, such that under this representation, point that are semantically closed together are closed together under the learnt metric space. In the work, the author fits a piece-wise linear embedding model on the data, and then propose the metric should follows several criterion based on this piece-wise linear embedding model. Lastly, the author train a neural network to transform the data, under the transformation, the L2 distance should satisfied the proposed criterion. The author assumes there exists a metric in the raw signal space that behaves locally consistent with the metric we want to learn. This allows us to find a neighborhood such that points in this neighborhood is semantically closed to each other. [I believe the author first use a pretrained neural network to extract feature from each image, then perform unsupervised deep metric learning on these features.] The author also assumes the neighborhood is locally low rank which can be approximated with a linear embedding.
Comments:
My background is in signal processing and unsupervised learning. I am not entirely update to date with the work on deep metric learning. My comments will be based on my summary. If anything is incorrect, please let me know, especially the part the bracket.
I think the proposed method seems to work well as it performs bette than existing benchmark. The resulting clustering from each piecewise linear manifold seems to perform better than other clustering algorithm. But I personally would like to see more qualitative and exploratory analysis than just the benchmark result. For example:
1. If the method is called piecewise-linear manifolds, can I visualize the piecewise-linear manifold structure? If the manifold of natural image is still very high dimensional to visualize, can the author to make a toy example for understanding and sanity check that the model indeed works as epected?
2. I wish I have a better understanding on what $o_{i,j}$ and $p_{i,j}$ is because they seems to be crucial. For my understanding, one is the distance between $x_j$ to $x_i$, the other one is the distance between $x_j$ to the manifold $x_i$ belongs to. Seem like these two distances needs to be balanced for the model to have good performance. But I don't get the intuition. The explanation in 4.4.3 still seem to obscure to me.
3. Is the local neighborhood indeed low-rank? The author seems to construct the neighbor by expanding if until it cannot be approximated by a linear subspace. I would like to see a more analysis on this. Like how much can you expand the neighborhood and how the eigenvalue of PCA fall off.
4. I think it would be nice if the author can compare their methods with other manifold learning method. The general assumption and goal is that the metric in the signal space works well locally but not globally. This seems to fit "think globally, fit locally" manifold learning very well.
5. In fact, I think the model can be consider a manifold learning model. The definition of the "Point-Point," "Proxy-Point" and "Proxy-Neighborhood" similarity together defined a similarity kernel, which is being approximated by the trained neural network. Does the author consider of not using the neural network? Since the signal space is already the feature of a pretrained neural network, maybe the similarity kernel can just be just approximated by a linear transformation. In this case, the model is a specific case of laplacian eigenmap. Or maybe the author should do some ablation study to see if a light weighted neural network will do a similar job, as this will fit the premise of the conference "Parsimony" better.
Saul, L and Roweis, S. Think globally, fit locally: unsupervised learning of low dimensional manifolds.
Tenenbaum, J. et al. A global geometric framework for nonlinear dimensionality reduction.
rating: 5: Marginally below acceptance threshold
confidence: 2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper |
Z0Fk5MyxkY | Piecewise-Linear Manifolds for Deep Metric Learning | [
"Shubhang Bhatnagar",
"Narendra Ahuja"
] | Unsupervised deep metric learning (UDML) focuses on learning a semantic representation space using only unlabeled data. This challenging problem requires accurately estimating the similarity between data points, which is used to supervise a deep network. For this purpose, we propose to model the high-dimensional data manifold using a piecewise-linear approximation, with each low-dimensional linear piece approximating the data manifold in a small neighborhood of a point. These neighborhoods are used to estimate similarity between data points. We empirically show that this similarity estimate correlates better with the ground truth than the similarity estimates of current state-of-the-art techniques. We also show that proxies, commonly used in supervised metric learning, can be used to model the piecewise-linear manifold in an unsupervised setting, helping improve performance. Our method outperforms existing unsupervised metric learning approaches on standard zero-shot image retrieval benchmarks. | [
"Deep metric learning",
"Unsupervised representation learning",
"Image retrieval"
] | https://openreview.net/pdf?id=Z0Fk5MyxkY | Ng1lPU7AuV | meta_review | 1,699,847,994,615 | Z0Fk5MyxkY | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission37/Area_Chair_Jxjx"
] | metareview: In this paper, the researchers delve into the problem of learning semantically meaningful representations in an unsupervised manner, particularly given the abundance of unlabeled data available. They focus on Unsupervised Deep Metric Learning (UDML), a field dedicated to harnessing this unlabeled data to acquire semantic similarity information. UDML typically relies on generating discrete pseudo-labels through clustering techniques, which can be noisy and inconsistent, posing a challenge in learning similarity effectively.
To address this issue, the researchers suggest modeling the data as a piecewise linear manifold, in contrast to treating it as a set of clusters. Each linear manifold approximates the local structure of data in a small neighborhood around a point. Linear manifolds offer greater label homogeneity compared to traditional clusters and enable the estimation of less noisy, continuous-valued similarity by measuring the fit error of the linear manifold to a point. Additionally, the researchers demonstrate that proxies, commonly employed in supervised metric learning, can also be utilized to model the piecewise linear manifold, even in an unsupervised setting. This inclusion contributes to improved performance and faster convergence.
The reviewers thought the paper was interesting and had good improvements. They raised various technical concerns many of which seem to have been addressed. Therefore I recommend acceptance but encourage the authors to address the remaining concerns in their final manuscript.
recommendation: Accept (Poster)
confidence: 4: The area chair is confident but not absolutely certain |
VP1Xrdz0Bp | HRBP: Hardware-friendly Regrouping towards Block-based Pruning for Sparse CNN Training | [
"Haoyu Ma",
"Chengming Zhang",
"lizhi xiang",
"Xiaolong Ma",
"Geng Yuan",
"Wenkai Zhang",
"Shiwei Liu",
"Tianlong Chen",
"Dingwen Tao",
"Yanzhi Wang",
"Zhangyang Wang",
"Xiaohui Xie"
] | Pruning at initialization and training a sparse network from scratch (sparse training) become increasingly popular. However, most sparse training literature addresses only the unstructured sparsity, which in practice brings little benefit to the training acceleration on GPU due to the irregularity of non-zero weights. In this paper, we work on sparse training with fine-grained structured sparsity, by extracting a few dense blocks from unstructured sparse weights. For Convolutional Neural networks (CNN), however, the extracted dense blocks will be broken in backpropagation due to the shape transformation of convolution filters implemented by GEMM. Thus, previous block-wise pruning methods can only be used to accelerate the forward pass of sparse CNN training. To this end, we propose Hardware-friendly Regrouping towards Block-based Pruning (HRBP), where the grouping is conducted on the kernel-wise mask. With HRBP, extracted dense blocks are preserved in backpropagation. Extensive experiments on CIFAR-10, CIFAR-100, and ImageNet demonstrate that HRBP can almost match the accuracy of unstructured sparse training methods while achieving a huge acceleration on hardware.
Code is available at: https://github.com/HowieMa/HRBP-pruning. | [
"efficient training",
"sparse training",
"fine-grained structured sparsity",
"regrouping algorithm"
] | https://openreview.net/pdf?id=VP1Xrdz0Bp | rlWWe1WTMe | official_review | 1,697,029,222,277 | VP1Xrdz0Bp | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission11/Reviewer_2ePh"
] | title: Official review
review: ## Summary:
The paper introduces HRBP, a block-wise pruning method for CNNs. HRBP maintains block-wise sparsity in both the forward and backward passes of CNN training, leading to improved efficiency.
## Strengths:
The idea of preserving block-wise sparsity in the backward pass is well-motivated.
Empirical results consistently demonstrate HRBP's performance improvements in static and dynamic training.
## Weaknesses & questions from the reviewer:
- To me, the technical exposition in the paper is quite challenging to follow. To clarify the core idea of HRBP, would it be possible to include a simple walk-through example, such as applying HRBP to a 1D convolution? This would demonstrate how HRBP maintains sparsity in both forward and backward passes in circulant matrices.
- In comparison to traditional block-pruning methods like [21], HRBP preserves sparsity in the backward pass. This is expected to enhance gradient computation efficiency, resulting in shorter run times. However, the experimental results also indicate that HRBP improves training accuracy compared to baseline methods. It would be insightful if the authors could provide some intuition behind this observation.
rating: 6: Marginally above acceptance threshold
confidence: 3: The reviewer is fairly confident that the evaluation is correct |
VP1Xrdz0Bp | HRBP: Hardware-friendly Regrouping towards Block-based Pruning for Sparse CNN Training | [
"Haoyu Ma",
"Chengming Zhang",
"lizhi xiang",
"Xiaolong Ma",
"Geng Yuan",
"Wenkai Zhang",
"Shiwei Liu",
"Tianlong Chen",
"Dingwen Tao",
"Yanzhi Wang",
"Zhangyang Wang",
"Xiaohui Xie"
] | Pruning at initialization and training a sparse network from scratch (sparse training) become increasingly popular. However, most sparse training literature addresses only the unstructured sparsity, which in practice brings little benefit to the training acceleration on GPU due to the irregularity of non-zero weights. In this paper, we work on sparse training with fine-grained structured sparsity, by extracting a few dense blocks from unstructured sparse weights. For Convolutional Neural networks (CNN), however, the extracted dense blocks will be broken in backpropagation due to the shape transformation of convolution filters implemented by GEMM. Thus, previous block-wise pruning methods can only be used to accelerate the forward pass of sparse CNN training. To this end, we propose Hardware-friendly Regrouping towards Block-based Pruning (HRBP), where the grouping is conducted on the kernel-wise mask. With HRBP, extracted dense blocks are preserved in backpropagation. Extensive experiments on CIFAR-10, CIFAR-100, and ImageNet demonstrate that HRBP can almost match the accuracy of unstructured sparse training methods while achieving a huge acceleration on hardware.
Code is available at: https://github.com/HowieMa/HRBP-pruning. | [
"efficient training",
"sparse training",
"fine-grained structured sparsity",
"regrouping algorithm"
] | https://openreview.net/pdf?id=VP1Xrdz0Bp | oD82Kiamwm | meta_review | 1,699,949,216,148 | VP1Xrdz0Bp | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission11/Area_Chair_BUUv"
] | metareview: Sparse DNNs are a prominent topic both in practice and in theory of deep learning but there are currently few off-the-shelf, simple ways to exploit weight sparsity. There is little low-level work which demonstrates important improvements with real data on real hardware.
Here we have a refreshingly sober paper which addresses this issue and designs a simple, creative solution which is demonstrated to work well in practice, with speed-ups on par with structured pruning methods and accuracy on par with unstructured methods. This could have substantial impact in practice. All reviewers agree that this is a solid contribution and recommend acceptance (either in score or in comments).
recommendation: Accept (Poster)
confidence: 5: The area chair is absolutely certain |
VP1Xrdz0Bp | HRBP: Hardware-friendly Regrouping towards Block-based Pruning for Sparse CNN Training | [
"Haoyu Ma",
"Chengming Zhang",
"lizhi xiang",
"Xiaolong Ma",
"Geng Yuan",
"Wenkai Zhang",
"Shiwei Liu",
"Tianlong Chen",
"Dingwen Tao",
"Yanzhi Wang",
"Zhangyang Wang",
"Xiaohui Xie"
] | Pruning at initialization and training a sparse network from scratch (sparse training) become increasingly popular. However, most sparse training literature addresses only the unstructured sparsity, which in practice brings little benefit to the training acceleration on GPU due to the irregularity of non-zero weights. In this paper, we work on sparse training with fine-grained structured sparsity, by extracting a few dense blocks from unstructured sparse weights. For Convolutional Neural networks (CNN), however, the extracted dense blocks will be broken in backpropagation due to the shape transformation of convolution filters implemented by GEMM. Thus, previous block-wise pruning methods can only be used to accelerate the forward pass of sparse CNN training. To this end, we propose Hardware-friendly Regrouping towards Block-based Pruning (HRBP), where the grouping is conducted on the kernel-wise mask. With HRBP, extracted dense blocks are preserved in backpropagation. Extensive experiments on CIFAR-10, CIFAR-100, and ImageNet demonstrate that HRBP can almost match the accuracy of unstructured sparse training methods while achieving a huge acceleration on hardware.
Code is available at: https://github.com/HowieMa/HRBP-pruning. | [
"efficient training",
"sparse training",
"fine-grained structured sparsity",
"regrouping algorithm"
] | https://openreview.net/pdf?id=VP1Xrdz0Bp | fQfzQk9mAh | decision | 1,700,497,692,121 | VP1Xrdz0Bp | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Program_Chairs"
] | decision: Accept (Oral)
comment: Overall, the reviewers and AC agreed that the paper makes a significant contribution to the field of sparse training. The paper proposes a novel hardware-friendly block-wise pruning method for sparse CNN training. The method preserves block-wise sparsity in both the forward and backward passes of CNN training, leading to improved efficiency. The paper is well-written and the experimental results are convincing.
The action PC chair for this paper is Gintare Karolina Dziugaite, who made the decision after carefully reading the paper as well as the comments by all reviewers and AC. The decision is agreed by all PC chairs.
title: Paper Decision |
VP1Xrdz0Bp | HRBP: Hardware-friendly Regrouping towards Block-based Pruning for Sparse CNN Training | [
"Haoyu Ma",
"Chengming Zhang",
"lizhi xiang",
"Xiaolong Ma",
"Geng Yuan",
"Wenkai Zhang",
"Shiwei Liu",
"Tianlong Chen",
"Dingwen Tao",
"Yanzhi Wang",
"Zhangyang Wang",
"Xiaohui Xie"
] | Pruning at initialization and training a sparse network from scratch (sparse training) become increasingly popular. However, most sparse training literature addresses only the unstructured sparsity, which in practice brings little benefit to the training acceleration on GPU due to the irregularity of non-zero weights. In this paper, we work on sparse training with fine-grained structured sparsity, by extracting a few dense blocks from unstructured sparse weights. For Convolutional Neural networks (CNN), however, the extracted dense blocks will be broken in backpropagation due to the shape transformation of convolution filters implemented by GEMM. Thus, previous block-wise pruning methods can only be used to accelerate the forward pass of sparse CNN training. To this end, we propose Hardware-friendly Regrouping towards Block-based Pruning (HRBP), where the grouping is conducted on the kernel-wise mask. With HRBP, extracted dense blocks are preserved in backpropagation. Extensive experiments on CIFAR-10, CIFAR-100, and ImageNet demonstrate that HRBP can almost match the accuracy of unstructured sparse training methods while achieving a huge acceleration on hardware.
Code is available at: https://github.com/HowieMa/HRBP-pruning. | [
"efficient training",
"sparse training",
"fine-grained structured sparsity",
"regrouping algorithm"
] | https://openreview.net/pdf?id=VP1Xrdz0Bp | eYwtSGsKvP | official_review | 1,696,305,390,278 | VP1Xrdz0Bp | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission11/Reviewer_gs6o"
] | title: Strong motivation and clear methodology, but the evaluation results are incomplete.
review: ### Strength:
* The paper clearly proposes a significant concern: the existing fine-grained structured pruning algorithm struggles to accelerate backpropagation when used with specific scenarios, such as CNN with GEMM. Section 3 delves deeply into this issue and presents a compelling argument.
* The method proposed focuses on regrouping using a kernel-wise mask and integrates pattern finding. This seems to be a sound approach to tackling the identified issues.
### Weakness:
* The presented results are incomplete and lack organization. In Table 1, for instance, the speedup values (which should be indicated in brackets as said in the caption) are missing. To grasp the benefits of the proposed method, readers have to cross-reference both Figure 5 and Figure 6, as these figures jointly demonstrate that the method offers improved speed without compromising performance. Notably, Figure 6 omits the ResNet-56 configuration, making it inconsistent with Figure 5.
* The paper does not provide sufficient details about its evaluation. The procedure by which training acceleration was gauged remains unclear, and the experimental setup is not adequately defined.
rating: 5: Marginally below acceptance threshold
confidence: 3: The reviewer is fairly confident that the evaluation is correct |
VP1Xrdz0Bp | HRBP: Hardware-friendly Regrouping towards Block-based Pruning for Sparse CNN Training | [
"Haoyu Ma",
"Chengming Zhang",
"lizhi xiang",
"Xiaolong Ma",
"Geng Yuan",
"Wenkai Zhang",
"Shiwei Liu",
"Tianlong Chen",
"Dingwen Tao",
"Yanzhi Wang",
"Zhangyang Wang",
"Xiaohui Xie"
] | Pruning at initialization and training a sparse network from scratch (sparse training) become increasingly popular. However, most sparse training literature addresses only the unstructured sparsity, which in practice brings little benefit to the training acceleration on GPU due to the irregularity of non-zero weights. In this paper, we work on sparse training with fine-grained structured sparsity, by extracting a few dense blocks from unstructured sparse weights. For Convolutional Neural networks (CNN), however, the extracted dense blocks will be broken in backpropagation due to the shape transformation of convolution filters implemented by GEMM. Thus, previous block-wise pruning methods can only be used to accelerate the forward pass of sparse CNN training. To this end, we propose Hardware-friendly Regrouping towards Block-based Pruning (HRBP), where the grouping is conducted on the kernel-wise mask. With HRBP, extracted dense blocks are preserved in backpropagation. Extensive experiments on CIFAR-10, CIFAR-100, and ImageNet demonstrate that HRBP can almost match the accuracy of unstructured sparse training methods while achieving a huge acceleration on hardware.
Code is available at: https://github.com/HowieMa/HRBP-pruning. | [
"efficient training",
"sparse training",
"fine-grained structured sparsity",
"regrouping algorithm"
] | https://openreview.net/pdf?id=VP1Xrdz0Bp | eYKFZyBz94 | official_review | 1,696,406,994,035 | VP1Xrdz0Bp | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission11/Reviewer_Turq"
] | title: Official Review of Reviewer Turq
review: This paper investigates practical hardware acceleration upon unstructured sparse training. Specifically, the authors offered a comprehensive analysis of the acceleration bottlenecks encountered by previous methods during backpropagation and then proposed a kernel-wise mask for grouping unstructured sparse weights, achieving effective acceleration during backpropagation. The efficacy of the proposed method is demonstrated through experiments on CIFAR and ImageNet. The reviewer acknowledges the contribution of this paper and also makes some suggestions as follows:
1. The author's analysis of GEMM is very thorough, which is highly appreciated. Nonetheless, I would like to raise two points. First, NVIDIA's sparse tensor core is implemented based on NHWC[1], so the T-mask itself devised for N:M pattern is acceptable. Moreover, the N:M sparsity of the network will not be applied to the situation where C0=3, which has been given in Nvidia's documentation, because it is inherently unsuitable. Secondly, I would like to point out that the authors' mentioned inapplicability of T-mask under NCHW situation can be overcome using a different, recently proposed BI-Mask[2]. Including a discussion around this might make the paper more detailed.
2. The kernel-wise mask is very innovative and easily understood, and its design thought that integrates hardware design is quite reasonable. It would be a significant contribution to the field if the author could open-source the related acceleration code.
3. On ImageNet, the authors mostly compared PAI methods. It would be better to compare with DST methods because accelerating training on ImageNet is more critical than smaller datasets like CIFAR. Even if the performance is not as good as DST, it can still show the user a trade-off between training acceleration and performance.
[1] Nvidia a100 tensor core gpu architecture. https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/nvidia-ampere-architecture-whitepaper.pdf, 2020
[2] Bi-directional Masks for Efficient N: M Sparse Training. In ICML, 2023.
rating: 7: Good paper, accept
confidence: 4: The reviewer is confident but not absolutely certain that the evaluation is correct |
VP1Xrdz0Bp | HRBP: Hardware-friendly Regrouping towards Block-based Pruning for Sparse CNN Training | [
"Haoyu Ma",
"Chengming Zhang",
"lizhi xiang",
"Xiaolong Ma",
"Geng Yuan",
"Wenkai Zhang",
"Shiwei Liu",
"Tianlong Chen",
"Dingwen Tao",
"Yanzhi Wang",
"Zhangyang Wang",
"Xiaohui Xie"
] | Pruning at initialization and training a sparse network from scratch (sparse training) become increasingly popular. However, most sparse training literature addresses only the unstructured sparsity, which in practice brings little benefit to the training acceleration on GPU due to the irregularity of non-zero weights. In this paper, we work on sparse training with fine-grained structured sparsity, by extracting a few dense blocks from unstructured sparse weights. For Convolutional Neural networks (CNN), however, the extracted dense blocks will be broken in backpropagation due to the shape transformation of convolution filters implemented by GEMM. Thus, previous block-wise pruning methods can only be used to accelerate the forward pass of sparse CNN training. To this end, we propose Hardware-friendly Regrouping towards Block-based Pruning (HRBP), where the grouping is conducted on the kernel-wise mask. With HRBP, extracted dense blocks are preserved in backpropagation. Extensive experiments on CIFAR-10, CIFAR-100, and ImageNet demonstrate that HRBP can almost match the accuracy of unstructured sparse training methods while achieving a huge acceleration on hardware.
Code is available at: https://github.com/HowieMa/HRBP-pruning. | [
"efficient training",
"sparse training",
"fine-grained structured sparsity",
"regrouping algorithm"
] | https://openreview.net/pdf?id=VP1Xrdz0Bp | Pyg3O49wMN | official_review | 1,697,412,800,101 | VP1Xrdz0Bp | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission11/Reviewer_UHbo"
] | title: Hardware friendly pruning for accelerating forward AND backward passes
review: This paper presents a pruning methodology which yields speed-ups of sparse neural networks in both forward and backward passes. The key intuition comes from how forward / backward passes are calculated in GEMM. As I understand it, the method prunes weights at the level of a single weight kernel, as indexed by input AND output channel. This guarantees that sparsity can be exploited to accelerate both forward and backward passes. This idea seems quite intuitive and simple. The results section shows that this pruning approach yields speed-ups on par with structured (channel-wise) pruning while exhibiting the accuracy of unstructured pruning. Authors also show results for dynamic sparsity training, which is an added plus. I rate this paper quite highly, as it shows real speed-ups on real hardware used by a large number of practitioners / researchers today.
rating: 8: Top 50% of accepted papers, clear accept
confidence: 4: The reviewer is confident but not absolutely certain that the evaluation is correct |
VP1Xrdz0Bp | HRBP: Hardware-friendly Regrouping towards Block-based Pruning for Sparse CNN Training | [
"Haoyu Ma",
"Chengming Zhang",
"lizhi xiang",
"Xiaolong Ma",
"Geng Yuan",
"Wenkai Zhang",
"Shiwei Liu",
"Tianlong Chen",
"Dingwen Tao",
"Yanzhi Wang",
"Zhangyang Wang",
"Xiaohui Xie"
] | Pruning at initialization and training a sparse network from scratch (sparse training) become increasingly popular. However, most sparse training literature addresses only the unstructured sparsity, which in practice brings little benefit to the training acceleration on GPU due to the irregularity of non-zero weights. In this paper, we work on sparse training with fine-grained structured sparsity, by extracting a few dense blocks from unstructured sparse weights. For Convolutional Neural networks (CNN), however, the extracted dense blocks will be broken in backpropagation due to the shape transformation of convolution filters implemented by GEMM. Thus, previous block-wise pruning methods can only be used to accelerate the forward pass of sparse CNN training. To this end, we propose Hardware-friendly Regrouping towards Block-based Pruning (HRBP), where the grouping is conducted on the kernel-wise mask. With HRBP, extracted dense blocks are preserved in backpropagation. Extensive experiments on CIFAR-10, CIFAR-100, and ImageNet demonstrate that HRBP can almost match the accuracy of unstructured sparse training methods while achieving a huge acceleration on hardware.
Code is available at: https://github.com/HowieMa/HRBP-pruning. | [
"efficient training",
"sparse training",
"fine-grained structured sparsity",
"regrouping algorithm"
] | https://openreview.net/pdf?id=VP1Xrdz0Bp | K8rHGTDqEe | meta_review | 1,700,394,110,206 | VP1Xrdz0Bp | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Program_Chairs"
] | metareview: Overall, the reviewers and AC agreed that the paper makes a significant contribution to the field of sparse training. The paper proposes a novel hardware-friendly block-wise pruning method for sparse CNN training. The method preserves block-wise sparsity in both the forward and backward passes of CNN training, leading to improved efficiency. The paper is well-written and the experimental results are convincing.
The action PC chair for this paper is Gintare Karolina Dziugaite, who made the decision after carefully reading the paper as well as the comments by all reviewers and AC. The decision is agreed by all PC chairs.
recommendation: Accept (Oral)
confidence: 5: The area chair is absolutely certain |
V7mcfiSjIT | Cross-Quality Few-Shot Transfer for Alloy Yield Strength Prediction: A New Materials Science Benchmark and A Sparsity-Oriented Optimization Framework | [
"Xuxi Chen",
"Tianlong Chen",
"Everardo Yeriel Olivares",
"Kate Elder",
"Scott McCall",
"Aurelien Perron",
"Joseph McKeown",
"Bhavya Kailkhura",
"Zhangyang Wang",
"Brian Gallagher"
] | Discovering high-entropy alloys (HEAs) with high yield strength (YS) is crucial in materials science. However, the YS can only be accurately measured by expensive and time-consuming experiments, hence cannot be acquired at scale. Learning-based methods could facilitate the discovery, but the lack of a comprehensive dataset on HEA YS has created barriers. We present X-Yield, a materials science benchmark with 240 experimentally measured (high-quality) and over 100,000 simulated (low-quality) HEA YS data. Due to the scarcity of experimental results and the quality gap with simulated data, existing transfer learning methods cannot generalize well on our dataset. We address this cross-quality few-shot transfer problem by leveraging model sparsification "twice" --- as a noise-robust feature regularizer at the pre-training stage, and as a data-efficient regularizer at the transfer stage. While the workflow already performs decently with sparsity patterns tuned independently for either stage, we propose a bi-level optimization framework termed Bi-RPT, that jointly learns optimal masks and allocates sparsity for both stages. The effectiveness of Bi-RPT is validated through experiments on X-Yield, alongside other testbeds. Specifically, we achieve a reduction of 8.9-19.8% in test MSE and a gain of 0.98-1.53% in test accuracy, using only 5-10% of the hard-to-generate real experimental data. The codes are available in https://github.com/VITA-Group/Bi-RPT. | [
"AI4Science",
"sparsity",
"bi-level optimization"
] | https://openreview.net/pdf?id=V7mcfiSjIT | UQcvoLnOCt | official_review | 1,697,418,814,394 | V7mcfiSjIT | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission20/Reviewer_6ADX"
] | title: Interesting topic, good writing, may need more comparison as a ML paper
review: The paper is motivated by the time consuming nature of discovering high entropy alloys (HEAs) with high yield strength (YS). Based on the difficulty, the authors try to use ML tools to estimate the YS of alloys directly. To prepare the training dataset, the authors combine a large quantity of simulation data and scarce real data. The authors then propose a novel approach, termed Bi-RPT to deal with the domain gap between the simulation data and real data. Empirical results validate the effectiveness of the proposed method over several baselines.
### Strength
The paper is well-motivated and well-written, and the proposed methods are introduced in detail with derivations.
Both the problem setting and the proposed method are interesting and novel at least to the reviewer's concern
### Weaknesses
The reviewer believes the paper in broad falls within the few-shot learning domain which is a well-estabilished field with plenty of prior work , and since this is an ML conference, when proposing a method, one application may not be enough to demonstrate the universal applicability. Hence the reviewer would suggest comparing the proposed method with more prior work and conducting the experiments on more datasets beyond the YS of alloy.
rating: 6: Marginally above acceptance threshold
confidence: 3: The reviewer is fairly confident that the evaluation is correct |
V7mcfiSjIT | Cross-Quality Few-Shot Transfer for Alloy Yield Strength Prediction: A New Materials Science Benchmark and A Sparsity-Oriented Optimization Framework | [
"Xuxi Chen",
"Tianlong Chen",
"Everardo Yeriel Olivares",
"Kate Elder",
"Scott McCall",
"Aurelien Perron",
"Joseph McKeown",
"Bhavya Kailkhura",
"Zhangyang Wang",
"Brian Gallagher"
] | Discovering high-entropy alloys (HEAs) with high yield strength (YS) is crucial in materials science. However, the YS can only be accurately measured by expensive and time-consuming experiments, hence cannot be acquired at scale. Learning-based methods could facilitate the discovery, but the lack of a comprehensive dataset on HEA YS has created barriers. We present X-Yield, a materials science benchmark with 240 experimentally measured (high-quality) and over 100,000 simulated (low-quality) HEA YS data. Due to the scarcity of experimental results and the quality gap with simulated data, existing transfer learning methods cannot generalize well on our dataset. We address this cross-quality few-shot transfer problem by leveraging model sparsification "twice" --- as a noise-robust feature regularizer at the pre-training stage, and as a data-efficient regularizer at the transfer stage. While the workflow already performs decently with sparsity patterns tuned independently for either stage, we propose a bi-level optimization framework termed Bi-RPT, that jointly learns optimal masks and allocates sparsity for both stages. The effectiveness of Bi-RPT is validated through experiments on X-Yield, alongside other testbeds. Specifically, we achieve a reduction of 8.9-19.8% in test MSE and a gain of 0.98-1.53% in test accuracy, using only 5-10% of the hard-to-generate real experimental data. The codes are available in https://github.com/VITA-Group/Bi-RPT. | [
"AI4Science",
"sparsity",
"bi-level optimization"
] | https://openreview.net/pdf?id=V7mcfiSjIT | MU9OSbkxRR | official_review | 1,696,648,912,227 | V7mcfiSjIT | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission20/Reviewer_73Cr"
] | title: This paper studies the discovery of high-entropy alloys (HEAs) with high yield strength (YS) in the material science with transfer learning. Because of the lack a comprehensive dataset and the time-consuming process for experimental data, the authors construct a new X-Y field dataset that consists of the majority of simulated data and minority of experimental data. Moreover, since the traditional transfer learning methods does not perform well on these datasets, the authors propose a sparsity regularizer for both pretraining stage on simulated data and transferring stage on experimental data. Furthermore, the authors incorporate a bi-level optimization framework to jointly learn the optimal masks and the allocation of sparsity for both stages.
review: Pros:
1. Applying the machine learning method to the material science is an interesting and important area.
2. A new benchmark for refractory HEA yield strength prediction is schedule to be public, which may inspire the investigation of new ML technology in the material science field.
3. The bi-level framework has a solid theoretical foundation.
4. This paper is well-written.
Cons:
1. the motivation of incorporation of sparsity is unclear. Although the authors claim that the use of sparse regularizers enhances the robustness and transferability, there is no related previous works in the line 198-202 to support such claim. Moreover, the incorporation of sparsity/pruning reduce the power of the networks, why it will be helpful for the performance in turn? Is this because the networks are easy to overfit on the simulated data?
In general, this paper is well-written and logically structured. The datasets may fill a missing gap in the material science domain and the sparsity regularizer with bi-level optimization is novel and solid. Therefore, I think it is qualified for the acceptance.
rating: 7: Good paper, accept
confidence: 4: The reviewer is confident but not absolutely certain that the evaluation is correct |
V7mcfiSjIT | Cross-Quality Few-Shot Transfer for Alloy Yield Strength Prediction: A New Materials Science Benchmark and A Sparsity-Oriented Optimization Framework | [
"Xuxi Chen",
"Tianlong Chen",
"Everardo Yeriel Olivares",
"Kate Elder",
"Scott McCall",
"Aurelien Perron",
"Joseph McKeown",
"Bhavya Kailkhura",
"Zhangyang Wang",
"Brian Gallagher"
] | Discovering high-entropy alloys (HEAs) with high yield strength (YS) is crucial in materials science. However, the YS can only be accurately measured by expensive and time-consuming experiments, hence cannot be acquired at scale. Learning-based methods could facilitate the discovery, but the lack of a comprehensive dataset on HEA YS has created barriers. We present X-Yield, a materials science benchmark with 240 experimentally measured (high-quality) and over 100,000 simulated (low-quality) HEA YS data. Due to the scarcity of experimental results and the quality gap with simulated data, existing transfer learning methods cannot generalize well on our dataset. We address this cross-quality few-shot transfer problem by leveraging model sparsification "twice" --- as a noise-robust feature regularizer at the pre-training stage, and as a data-efficient regularizer at the transfer stage. While the workflow already performs decently with sparsity patterns tuned independently for either stage, we propose a bi-level optimization framework termed Bi-RPT, that jointly learns optimal masks and allocates sparsity for both stages. The effectiveness of Bi-RPT is validated through experiments on X-Yield, alongside other testbeds. Specifically, we achieve a reduction of 8.9-19.8% in test MSE and a gain of 0.98-1.53% in test accuracy, using only 5-10% of the hard-to-generate real experimental data. The codes are available in https://github.com/VITA-Group/Bi-RPT. | [
"AI4Science",
"sparsity",
"bi-level optimization"
] | https://openreview.net/pdf?id=V7mcfiSjIT | JRy2Elzf1q | meta_review | 1,700,014,657,489 | V7mcfiSjIT | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission20/Area_Chair_bojB"
] | metareview: This paper has made nice contributions for Alloy Yield Strength Prediction. First, it establishes a new benchmark, which fills the gap in existing literature and would be useful for future study. Second, it considers the domain gap between simulation data and experimental data and proposes a sparsity-oriented bilevel optimization methods to prune the network learned from the source data and also the one fine tuned on the target domain. All reviewers think the paper is well written. The authors have tried to address the concerns raised in comments. Based on these, I will recommend acceptance.
recommendation: Accept (Poster)
confidence: 5: The area chair is absolutely certain |
V7mcfiSjIT | Cross-Quality Few-Shot Transfer for Alloy Yield Strength Prediction: A New Materials Science Benchmark and A Sparsity-Oriented Optimization Framework | [
"Xuxi Chen",
"Tianlong Chen",
"Everardo Yeriel Olivares",
"Kate Elder",
"Scott McCall",
"Aurelien Perron",
"Joseph McKeown",
"Bhavya Kailkhura",
"Zhangyang Wang",
"Brian Gallagher"
] | Discovering high-entropy alloys (HEAs) with high yield strength (YS) is crucial in materials science. However, the YS can only be accurately measured by expensive and time-consuming experiments, hence cannot be acquired at scale. Learning-based methods could facilitate the discovery, but the lack of a comprehensive dataset on HEA YS has created barriers. We present X-Yield, a materials science benchmark with 240 experimentally measured (high-quality) and over 100,000 simulated (low-quality) HEA YS data. Due to the scarcity of experimental results and the quality gap with simulated data, existing transfer learning methods cannot generalize well on our dataset. We address this cross-quality few-shot transfer problem by leveraging model sparsification "twice" --- as a noise-robust feature regularizer at the pre-training stage, and as a data-efficient regularizer at the transfer stage. While the workflow already performs decently with sparsity patterns tuned independently for either stage, we propose a bi-level optimization framework termed Bi-RPT, that jointly learns optimal masks and allocates sparsity for both stages. The effectiveness of Bi-RPT is validated through experiments on X-Yield, alongside other testbeds. Specifically, we achieve a reduction of 8.9-19.8% in test MSE and a gain of 0.98-1.53% in test accuracy, using only 5-10% of the hard-to-generate real experimental data. The codes are available in https://github.com/VITA-Group/Bi-RPT. | [
"AI4Science",
"sparsity",
"bi-level optimization"
] | https://openreview.net/pdf?id=V7mcfiSjIT | JAFH3qnPZQ | decision | 1,700,430,805,764 | V7mcfiSjIT | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Program_Chairs"
] | decision: Accept (Oral)
comment: All reviewers and AC agreed that the paper is of high quality. This paper has made nice contributions for Alloy Yield Strength Prediction. First, it establishes a new benchmark, which fills the gap in existing literature and would be useful for future study. Second, it considers the domain gap between simulation data and experimental data and proposes a sparsity-oriented bilevel optimization methods to prune the network learned from the source data and also the one fine tuned on the target domain.
The action PC chair for this paper is Qing Qu, who made the decision after carefully reading the paper as well as the comments by all reviewers and AC. The decision is agreed upon by all PC chairs.
title: Paper Decision |
V7mcfiSjIT | Cross-Quality Few-Shot Transfer for Alloy Yield Strength Prediction: A New Materials Science Benchmark and A Sparsity-Oriented Optimization Framework | [
"Xuxi Chen",
"Tianlong Chen",
"Everardo Yeriel Olivares",
"Kate Elder",
"Scott McCall",
"Aurelien Perron",
"Joseph McKeown",
"Bhavya Kailkhura",
"Zhangyang Wang",
"Brian Gallagher"
] | Discovering high-entropy alloys (HEAs) with high yield strength (YS) is crucial in materials science. However, the YS can only be accurately measured by expensive and time-consuming experiments, hence cannot be acquired at scale. Learning-based methods could facilitate the discovery, but the lack of a comprehensive dataset on HEA YS has created barriers. We present X-Yield, a materials science benchmark with 240 experimentally measured (high-quality) and over 100,000 simulated (low-quality) HEA YS data. Due to the scarcity of experimental results and the quality gap with simulated data, existing transfer learning methods cannot generalize well on our dataset. We address this cross-quality few-shot transfer problem by leveraging model sparsification "twice" --- as a noise-robust feature regularizer at the pre-training stage, and as a data-efficient regularizer at the transfer stage. While the workflow already performs decently with sparsity patterns tuned independently for either stage, we propose a bi-level optimization framework termed Bi-RPT, that jointly learns optimal masks and allocates sparsity for both stages. The effectiveness of Bi-RPT is validated through experiments on X-Yield, alongside other testbeds. Specifically, we achieve a reduction of 8.9-19.8% in test MSE and a gain of 0.98-1.53% in test accuracy, using only 5-10% of the hard-to-generate real experimental data. The codes are available in https://github.com/VITA-Group/Bi-RPT. | [
"AI4Science",
"sparsity",
"bi-level optimization"
] | https://openreview.net/pdf?id=V7mcfiSjIT | AfeepDqCJs | official_review | 1,696,434,384,312 | V7mcfiSjIT | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission20/Reviewer_Kp6T"
] | title: Interesting, but needs a few more baselines
review: This paper considers the very challenging scenario in material science where one hopes to obtain a predictor of yield strength (YS) of high-entropy alloys (HEAs) when high quality real-world measurements are very scarce and simulation results are not very accurate.
The paper makes 2 main contributions:
1. The authors compiled a new benchmark dataset “X-Yield” that represents the scenario described above.
2. The authors proposed a new sparsity-regularized training scheme (Bi-RPT) which achieves better results than plain pre-training under the setup of strong domain gap and huge dataset size difference between pre-training and fine-tuning.
Overall, the paper is well written and easy to follow. The experimental results are clear and convincing. I do have a few questions and suggestions to the authors which I will detail below:
It is interesting that the proposed Bi-RPT method can improve the transfer performance. My understanding is that this method outperforms baseline pretrain-and-transfer only under strong domain gap settings. From a scientific point of view, a proof-of-concept experiment needs to be included to show a setting where Bi-RPT doesn’t help. For example, when transferring from ImageNet to ImageNetV2, sparsity regularization would likely underperform plain pretrain-transfer. If otherwise, this method should be considered to be a better transfer learning method in general.
From the experimental results in Table 1, it seems that sparsity at the transfer stage doesn’t change the result significantly. Pre-training sparsity helps the result a little bit, and allowing optimization of pre-training sparsity masks on fine-tuning data provides additional gains. This still holds in Table 4,5,6,7 in supplementary materials, although sometimes tuning the transfer sparsity helps, the difference is very minor. Could it be possible that not using the transfer mask in Bi-RPT produces identical results and the method can thus be simplified？
When applying Bi-RPT to X-Yield dataset, the authors transformed the composition of the material and the temperature parameter to a pseudo-image, then used the same image classification Convnet to solve the problem. My concern is that this method is unnecessarily complex, given the material-science problem considered is mostly low-dimensional, could it be that classical regression method like LASSO, or simple MLP works equally well or even better? Since in that setting, similar sparsity constraints can be applied, and the number of parameters is much less. Either way, it would be nicer to include some baseline comparisons to simpler methods to show the necessity of employing a deep CNN.
I’ll update my score if the questions above are sufficiently addressed by the authors.
Disclaimer: I have no expertise in material science and none of my evaluation is based on the material science significance of this paper.
rating: 6
confidence: 4 |
S8MPHInGqj | Deep Leakage from Model in Federated Learning | [
"Zihao Zhao",
"Mengen Luo",
"Wenbo Ding"
] | Federated Learning (FL) was conceived as a secure form of distributed learning by keeping private training data local and only communicating public model gradients between clients. However, a slew of gradient leakage attacks proposed to date undermine this claim by proving its insecurity. A common limitation of these attacks is the necessity for extensive auxiliary information, such as model weights, optimizers, and certain hyperparameters (e.g., learning rate), which are challenging to acquire in practical scenarios. Furthermore, several existing algorithms, including FedAvg, circumvent the transmission of model gradients in FL by instead sending model weights, but the potential security breaches of this approach are seldom considered. In this paper, we propose two innovative frameworks, DLM and DLM+, that reveal the potential leakage of private local data of clients when transmitting model weights under the FL framework. We also conduct a series of experiments to elucidate the impact and universality of our attack frameworks. Additionally, we propose and evaluate two defenses against the proposed attacks, assessing their protective efficacy. | [
"Federated learning",
"distributed learning",
"privacy leakage"
] | https://openreview.net/pdf?id=S8MPHInGqj | zjPM7V9jIV | decision | 1,700,422,533,506 | S8MPHInGqj | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Program_Chairs"
] | decision: Accept (Oral)
comment: This paper develops new attack algorithms, from which private local data might be leaked when model weights are transmitted in the FL framework. Extensive experiments are provided to corroborate the effectiveness of the proposed approaches.
The action PC chair for this paper is Yuejie Chi, who made the decision after carefully reading the paper as well as the comments by all reviewers and AC. The decision is agreed by all PC chairs.
title: Paper Decision |
S8MPHInGqj | Deep Leakage from Model in Federated Learning | [
"Zihao Zhao",
"Mengen Luo",
"Wenbo Ding"
] | Federated Learning (FL) was conceived as a secure form of distributed learning by keeping private training data local and only communicating public model gradients between clients. However, a slew of gradient leakage attacks proposed to date undermine this claim by proving its insecurity. A common limitation of these attacks is the necessity for extensive auxiliary information, such as model weights, optimizers, and certain hyperparameters (e.g., learning rate), which are challenging to acquire in practical scenarios. Furthermore, several existing algorithms, including FedAvg, circumvent the transmission of model gradients in FL by instead sending model weights, but the potential security breaches of this approach are seldom considered. In this paper, we propose two innovative frameworks, DLM and DLM+, that reveal the potential leakage of private local data of clients when transmitting model weights under the FL framework. We also conduct a series of experiments to elucidate the impact and universality of our attack frameworks. Additionally, we propose and evaluate two defenses against the proposed attacks, assessing their protective efficacy. | [
"Federated learning",
"distributed learning",
"privacy leakage"
] | https://openreview.net/pdf?id=S8MPHInGqj | t1T6zI9FhW | meta_review | 1,699,851,660,787 | S8MPHInGqj | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission16/Area_Chair_rwYk"
] | metareview: This paper develops a new attack algorithm for model inversion attacks in federated learning. Different from the majority of prior art, this paper considers the setting where models instead of gradients are being transmitted. All three reviewers agree that this is a novel and interesting work and deserves acceptance to the CPAL conference. During the rebuttal phase, the authors further address the reviewers' minor concerns on this paper's complexity and setting. Therefore, I support the acceptance of this paper.
recommendation: Accept (Poster)
confidence: 5: The area chair is absolutely certain |
S8MPHInGqj | Deep Leakage from Model in Federated Learning | [
"Zihao Zhao",
"Mengen Luo",
"Wenbo Ding"
] | Federated Learning (FL) was conceived as a secure form of distributed learning by keeping private training data local and only communicating public model gradients between clients. However, a slew of gradient leakage attacks proposed to date undermine this claim by proving its insecurity. A common limitation of these attacks is the necessity for extensive auxiliary information, such as model weights, optimizers, and certain hyperparameters (e.g., learning rate), which are challenging to acquire in practical scenarios. Furthermore, several existing algorithms, including FedAvg, circumvent the transmission of model gradients in FL by instead sending model weights, but the potential security breaches of this approach are seldom considered. In this paper, we propose two innovative frameworks, DLM and DLM+, that reveal the potential leakage of private local data of clients when transmitting model weights under the FL framework. We also conduct a series of experiments to elucidate the impact and universality of our attack frameworks. Additionally, we propose and evaluate two defenses against the proposed attacks, assessing their protective efficacy. | [
"Federated learning",
"distributed learning",
"privacy leakage"
] | https://openreview.net/pdf?id=S8MPHInGqj | kB6AgQ0m5b | official_review | 1,696,698,240,536 | S8MPHInGqj | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission16/Reviewer_GcxC"
] | title: Official Review of Submission 16
review: ## Overview
The study introduces a novel attack framework, termed DLM and DLM+, that challenges existing federated learning systems where model weights are transmitted rather than the gradient. Unique to DLM(+) is its reliance solely on communicated model parameters and the loss function, making it more flexible than earlier gradient leakage attacks. Comprehensive tests underscore the consistent efficacy of the introduced methods.
## Comments
(+) The advancements brought about by DLM+ are notably superior to other attack baselines.
(+) A series of ablation studies have been carried out, which strongly attest to the efficacy of the proposed method.
(+) This work is well-motivated and proposes a threat to federated learning where only model weights are transmitted.
(-) Referring to line 172, while MNIST, CIFAR10, and CIFAR100 are employed to encompass different image sizes, both sets predominantly contain smaller images (28x28 for MNIST and 32x32 for CIFAR10/100). Integrating datasets with larger images might offer more comprehensive results.
(-) It would be valuable to understand if the presented method can be extended to situations where prior information about the loss function isn't accessible.
rating: 6: Marginally above acceptance threshold
confidence: 3: The reviewer is fairly confident that the evaluation is correct |
S8MPHInGqj | Deep Leakage from Model in Federated Learning | [
"Zihao Zhao",
"Mengen Luo",
"Wenbo Ding"
] | Federated Learning (FL) was conceived as a secure form of distributed learning by keeping private training data local and only communicating public model gradients between clients. However, a slew of gradient leakage attacks proposed to date undermine this claim by proving its insecurity. A common limitation of these attacks is the necessity for extensive auxiliary information, such as model weights, optimizers, and certain hyperparameters (e.g., learning rate), which are challenging to acquire in practical scenarios. Furthermore, several existing algorithms, including FedAvg, circumvent the transmission of model gradients in FL by instead sending model weights, but the potential security breaches of this approach are seldom considered. In this paper, we propose two innovative frameworks, DLM and DLM+, that reveal the potential leakage of private local data of clients when transmitting model weights under the FL framework. We also conduct a series of experiments to elucidate the impact and universality of our attack frameworks. Additionally, we propose and evaluate two defenses against the proposed attacks, assessing their protective efficacy. | [
"Federated learning",
"distributed learning",
"privacy leakage"
] | https://openreview.net/pdf?id=S8MPHInGqj | EOewOuPcqd | official_review | 1,697,403,715,887 | S8MPHInGqj | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission16/Reviewer_Gbfu"
] | title: Review for deep leakage
review: 1. Summary:
The paper begins by highlighting the challenges posed by the growth and complexity of data to traditional centralized machine learning. Distributed learning has emerged as a solution, with federated learning (FL) being a notable application. FL aims to protect client data privacy by keeping training data local. However, even without uploading raw training data, FL can still be vulnerable to data leakage.
2. Main Contributions:
a) The paper identifies the possibility of recovering private training data in FL using only transmitted model parameters and loss functions, challenging the foundational security of FL.
b) Two novel attack frameworks, DLM and DLM+, are introduced. These frameworks are applied to FedAvg, a widely-used algorithm in FL. c) The results show that FL architectures that exchange model weights cannot fully protect client data.
d) The paper compares the proposed model leakage attacks with existing gradient leakage attacks, demonstrating the superiority of the new attacks. Additionally, two defenses against these attacks are introduced.
3. Pros:
a) The paper addresses a critical security concern in federated learning, enhancing the understanding of potential vulnerabilities.
b) Introduces two novel attack frameworks, DLM and DLM+, offering a comprehensive approach to understanding data leakage in FL.
c) Provides defenses against the proposed attacks, ensuring a balanced perspective on the issue.
4. Cons:
a) The paper's focus on transmitted model weights might limit its applicability to FL systems that don't rely on this method.
b) The complexity of the proposed frameworks might make them challenging to implement in real-world scenarios.
c) The paper assumes that attackers have access to certain transmitted data, which might not always be the case in secure FL implementations.
rating: 6: Marginally above acceptance threshold
confidence: 3: The reviewer is fairly confident that the evaluation is correct |
S8MPHInGqj | Deep Leakage from Model in Federated Learning | [
"Zihao Zhao",
"Mengen Luo",
"Wenbo Ding"
] | Federated Learning (FL) was conceived as a secure form of distributed learning by keeping private training data local and only communicating public model gradients between clients. However, a slew of gradient leakage attacks proposed to date undermine this claim by proving its insecurity. A common limitation of these attacks is the necessity for extensive auxiliary information, such as model weights, optimizers, and certain hyperparameters (e.g., learning rate), which are challenging to acquire in practical scenarios. Furthermore, several existing algorithms, including FedAvg, circumvent the transmission of model gradients in FL by instead sending model weights, but the potential security breaches of this approach are seldom considered. In this paper, we propose two innovative frameworks, DLM and DLM+, that reveal the potential leakage of private local data of clients when transmitting model weights under the FL framework. We also conduct a series of experiments to elucidate the impact and universality of our attack frameworks. Additionally, we propose and evaluate two defenses against the proposed attacks, assessing their protective efficacy. | [
"Federated learning",
"distributed learning",
"privacy leakage"
] | https://openreview.net/pdf?id=S8MPHInGqj | 6GTQrCBWLQ | official_review | 1,696,776,367,592 | S8MPHInGqj | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission16/Reviewer_wxwi"
] | title: Simple changes to DLG leads to higher attack performance and robustness
review: The paper titled "Deep Leakage from Model Parameters in Federated Learning" investigates the security vulnerabilities within the Federated Learning (FL) framework, primarily focusing on the potential data leakage when transmitting model parameters. The authors highlight the importance of FL in addressing data privacy concerns and the challenges posed by the explosive growth of data. The authors set the stage by mentioning the recent advancements in gradient leakage attacks, which have raised concerns about the security of FL. They emphasize the common requirement for auxiliary information in these attacks, such as model weights, optimizers, and hyperparameters, which can be challenging to obtain in practical scenarios. Furthermore, the paper raises concerns about the security of transmitting model weights in FL, a less explored area. To address these issues, the authors propose two novel attack frameworks, DLM and DLM+, which aim to expose the potential leakage of private client data when transmitting model weights.
To reduce the number of trained variables, the authors propose to approximate learning rate with value of the weight difference of global model divided by gradients. Doing so can get rid of learning rate and gain better performance. Empirical results verify the effectiveness of proposed method. In addition, two defenses against the proposed attacks have been presented.
Pros:
(1) The paper is written clearly and present the main idea and methodology in a very pleasant way. I can directly understand the main mechanism of data leakage after reading the paper once.
(2) The main idea is simple yet effective. A very straight-forward modification leads to significant improvement. I believe the simplicity is a pro rather than con in this case.
(3) Empirical results demonstrate the effectiveness of the proposed method clearly.
(4) The proposed method enjoys appealing robustness to local iteration and optimizer as well.
Cons:
(1) While promising performance is demonstrated, I would like to see the efficiency comparison between DLM+ and DLG. Does the simple change leads to higher computational cost?
rating: 7: Good paper, accept
confidence: 3: The reviewer is fairly confident that the evaluation is correct |
Pby7pqtBYN | Balance is Essence: Accelerating Sparse Training via Adaptive Gradient Correction | [
"Bowen Lei",
"Dongkuan Xu",
"Ruqi Zhang",
"Shuren He",
"Bani Mallick"
] | Despite impressive performance, deep neural networks require significant memory and computation costs, prohibiting their application in resource-constrained scenarios. Sparse training is one of the most common techniques to reduce these costs, however, the sparsity constraints add difficulty to the optimization, resulting in an increase in training time and instability. In this work, we aim to overcome this problem and achieve space-time co-efficiency. To accelerate and stabilize the convergence of sparse training, we analyze the gradient changes and develop an adaptive gradient correction method. Specifically, we approximate the correlation between the current and previous gradients, which is used to balance the two gradients to obtain a corrected gradient. Our method can be used with the most popular sparse training pipelines under both standard and adversarial setups. Theoretically, we prove that our method can accelerate the convergence rate of sparse training. Extensive experiments on multiple datasets, model architectures, and sparsities demonstrate that our method outperforms leading sparse training methods by up to \textbf{5.0\%} in accuracy given the same number of training epochs, and reduces the number of training epochs by up to \textbf{52.1\%} to achieve the same accuracy. Our code is available on: \url{https://github.com/StevenBoys/AGENT}. | [
"Sparse Training",
"Space-time Co-efficiency",
"Acceleration",
"Stability",
"Gradient Correction"
] | https://openreview.net/pdf?id=Pby7pqtBYN | zCHnZCsnYW | decision | 1,700,361,534,019 | Pby7pqtBYN | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Program_Chairs"
] | decision: Accept (Oral)
comment: The paper introduces a novel optimization algorithm for sparse neural network training, receiving positive feedback for its innovative approach and theoretical analysis. Reviewers appreciate the overall clear writing and structure but still suggest a few clarity improvements. While some valid concerns about computational intensity for large datasets are raised, the convincing results and unanimous recommendation for acceptance indicate strong support for the paper's contributions.
The action PC chair for this paper is Atlas Wang, who made the decision after carefully reading the paper as well as the comments by all reviewers and AC. The decision is agreed by all PC chairs.
title: Paper Decision |
Pby7pqtBYN | Balance is Essence: Accelerating Sparse Training via Adaptive Gradient Correction | [
"Bowen Lei",
"Dongkuan Xu",
"Ruqi Zhang",
"Shuren He",
"Bani Mallick"
] | Despite impressive performance, deep neural networks require significant memory and computation costs, prohibiting their application in resource-constrained scenarios. Sparse training is one of the most common techniques to reduce these costs, however, the sparsity constraints add difficulty to the optimization, resulting in an increase in training time and instability. In this work, we aim to overcome this problem and achieve space-time co-efficiency. To accelerate and stabilize the convergence of sparse training, we analyze the gradient changes and develop an adaptive gradient correction method. Specifically, we approximate the correlation between the current and previous gradients, which is used to balance the two gradients to obtain a corrected gradient. Our method can be used with the most popular sparse training pipelines under both standard and adversarial setups. Theoretically, we prove that our method can accelerate the convergence rate of sparse training. Extensive experiments on multiple datasets, model architectures, and sparsities demonstrate that our method outperforms leading sparse training methods by up to \textbf{5.0\%} in accuracy given the same number of training epochs, and reduces the number of training epochs by up to \textbf{52.1\%} to achieve the same accuracy. Our code is available on: \url{https://github.com/StevenBoys/AGENT}. | [
"Sparse Training",
"Space-time Co-efficiency",
"Acceleration",
"Stability",
"Gradient Correction"
] | https://openreview.net/pdf?id=Pby7pqtBYN | qBFQ5RXW4g | official_review | 1,696,921,486,680 | Pby7pqtBYN | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission17/Reviewer_cG53"
] | title: Review
review: Pros:
---
- I am a big fan of this idea. Correcting the gradient for sparse training is novel and well-motivated.
- The paper provides not only empirical results but also theoretical analysis.
- The writing is lucid, and the paper is well-structured.
- Though AGENT inevitably adds extra FLOPs, which seems contradictory to the efficiency goal of sparse training, the authors offer potential solutions to this concern.
Cons:
---
- Formatting requires further polishing, e.g., a missing period before "However" in line 84. Also, the first letter be capitalized in figure 1 but not in other figures.
- I am curious if AGENT also improves the gradient norm which is important for sparse training as mentioned in [1].
- It would be intriguing to see experiments conducted on larger scale models or datasets, for example, the ViT model on ImageNet.
[1] Gradient Flow in Sparse Neural Networks and How Lottery Tickets Win, AAAI. U Evci. et al.
rating: 7: Good paper, accept
confidence: 4: The reviewer is confident but not absolutely certain that the evaluation is correct |
Pby7pqtBYN | Balance is Essence: Accelerating Sparse Training via Adaptive Gradient Correction | [
"Bowen Lei",
"Dongkuan Xu",
"Ruqi Zhang",
"Shuren He",
"Bani Mallick"
] | Despite impressive performance, deep neural networks require significant memory and computation costs, prohibiting their application in resource-constrained scenarios. Sparse training is one of the most common techniques to reduce these costs, however, the sparsity constraints add difficulty to the optimization, resulting in an increase in training time and instability. In this work, we aim to overcome this problem and achieve space-time co-efficiency. To accelerate and stabilize the convergence of sparse training, we analyze the gradient changes and develop an adaptive gradient correction method. Specifically, we approximate the correlation between the current and previous gradients, which is used to balance the two gradients to obtain a corrected gradient. Our method can be used with the most popular sparse training pipelines under both standard and adversarial setups. Theoretically, we prove that our method can accelerate the convergence rate of sparse training. Extensive experiments on multiple datasets, model architectures, and sparsities demonstrate that our method outperforms leading sparse training methods by up to \textbf{5.0\%} in accuracy given the same number of training epochs, and reduces the number of training epochs by up to \textbf{52.1\%} to achieve the same accuracy. Our code is available on: \url{https://github.com/StevenBoys/AGENT}. | [
"Sparse Training",
"Space-time Co-efficiency",
"Acceleration",
"Stability",
"Gradient Correction"
] | https://openreview.net/pdf?id=Pby7pqtBYN | njGJXKnXgO | official_review | 1,697,410,376,672 | Pby7pqtBYN | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission17/Reviewer_muE8"
] | title: Adaptive gradient correction for sparse neural network training
review: This paper proposes a new optimization algorithm for sparse neural network training. The authors provide an intuitive explanation of the algorithm, empirical motivation by examining gradient variance and correlation as a function of sparsity, theoretical justification for their approach, and empirical results on cifar 10, cifar 100, and imagenet.
I was slightly confused about the complexity of the approach. In algorithm 1, it appears that every m steps, we must calculate g^tilde, which requires traversing the entire dataset to calculate the gradient at the current model weights. If I am understanding the algorithm correctly, this is a pretty computationally intensive requirement. Moreover, this may make this algorithm prohibitively expensive when 1) the dataset size is effectively infinite (i.e. streaming data applications which are common in industry - https://arxiv.org/pdf/1906.00091.pdf); 2) storing and accessing an extra set of gradients is expensive from the point of view of (GPU) memory (i.e. foundation models, large language models, recommendation models https://arxiv.org/pdf/2208.06399.pdf).
On the other hand, the results in section 6 are quite convincing. Since I am not an expert in the space of sparse training algorithms, I will give the authors the benefit of the doubt that their approach is scalable when giving my rating.
rating: 7: Good paper, accept
confidence: 3: The reviewer is fairly confident that the evaluation is correct |
Pby7pqtBYN | Balance is Essence: Accelerating Sparse Training via Adaptive Gradient Correction | [
"Bowen Lei",
"Dongkuan Xu",
"Ruqi Zhang",
"Shuren He",
"Bani Mallick"
] | Despite impressive performance, deep neural networks require significant memory and computation costs, prohibiting their application in resource-constrained scenarios. Sparse training is one of the most common techniques to reduce these costs, however, the sparsity constraints add difficulty to the optimization, resulting in an increase in training time and instability. In this work, we aim to overcome this problem and achieve space-time co-efficiency. To accelerate and stabilize the convergence of sparse training, we analyze the gradient changes and develop an adaptive gradient correction method. Specifically, we approximate the correlation between the current and previous gradients, which is used to balance the two gradients to obtain a corrected gradient. Our method can be used with the most popular sparse training pipelines under both standard and adversarial setups. Theoretically, we prove that our method can accelerate the convergence rate of sparse training. Extensive experiments on multiple datasets, model architectures, and sparsities demonstrate that our method outperforms leading sparse training methods by up to \textbf{5.0\%} in accuracy given the same number of training epochs, and reduces the number of training epochs by up to \textbf{52.1\%} to achieve the same accuracy. Our code is available on: \url{https://github.com/StevenBoys/AGENT}. | [
"Sparse Training",
"Space-time Co-efficiency",
"Acceleration",
"Stability",
"Gradient Correction"
] | https://openreview.net/pdf?id=Pby7pqtBYN | ZZTxpyLUgU | official_review | 1,697,204,053,264 | Pby7pqtBYN | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission17/Reviewer_XGVR"
] | title: Official Review of Paper17 by Reviewer xgvr
review: ### Summary
The paper introduces a novel method named AGENT for implementing the adaptive gradient correction technique to accelerate and stabilize the convergence of sparse training. The author provides a theoretical analysis and conducts extensive experiments, demonstrating up to a 5% improvement.
### Quality, clarity, originality
The algorithm is novel , and the code for reproducing the experiments is provided. The paper is well-written, clear, and comprehensive.
### Strength and weaknesses:
Strength:
+ A good level of novelty
+ Extensive experiments
+ Solid theoretical analysis
Weaknesses:
+ While convergence is faster and more stable, there is no improvement in final accuracy on large datasets.
rating: 9: Top 15% of accepted papers, strong accept
confidence: 5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature |
Pby7pqtBYN | Balance is Essence: Accelerating Sparse Training via Adaptive Gradient Correction | [
"Bowen Lei",
"Dongkuan Xu",
"Ruqi Zhang",
"Shuren He",
"Bani Mallick"
] | Despite impressive performance, deep neural networks require significant memory and computation costs, prohibiting their application in resource-constrained scenarios. Sparse training is one of the most common techniques to reduce these costs, however, the sparsity constraints add difficulty to the optimization, resulting in an increase in training time and instability. In this work, we aim to overcome this problem and achieve space-time co-efficiency. To accelerate and stabilize the convergence of sparse training, we analyze the gradient changes and develop an adaptive gradient correction method. Specifically, we approximate the correlation between the current and previous gradients, which is used to balance the two gradients to obtain a corrected gradient. Our method can be used with the most popular sparse training pipelines under both standard and adversarial setups. Theoretically, we prove that our method can accelerate the convergence rate of sparse training. Extensive experiments on multiple datasets, model architectures, and sparsities demonstrate that our method outperforms leading sparse training methods by up to \textbf{5.0\%} in accuracy given the same number of training epochs, and reduces the number of training epochs by up to \textbf{52.1\%} to achieve the same accuracy. Our code is available on: \url{https://github.com/StevenBoys/AGENT}. | [
"Sparse Training",
"Space-time Co-efficiency",
"Acceleration",
"Stability",
"Gradient Correction"
] | https://openreview.net/pdf?id=Pby7pqtBYN | YVJ45jebph | meta_review | 1,699,837,191,426 | Pby7pqtBYN | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission17/Area_Chair_uw8B"
] | metareview: This paper proposes an accelerated sparse training algorithm for neural networks, and provides a theoretical convergence analysis as well as extensive experiments. The reviewers all agree that this is a solid paper and unanimously recommend accept.
recommendation: Accept (Oral)
confidence: 4: The area chair is confident but not absolutely certain |
NFqCYfoLw6 | An Adaptive Tangent Feature Perspective of Neural Networks | [
"Daniel LeJeune",
"Sina Alemohammad"
] | In order to better understand feature learning in neural networks, we propose and study linear models in tangent feature space where the features are allowed to be transformed during training. We consider linear feature transformations, resulting in a joint optimization over parameters and transformations with a bilinear interpolation constraint. We show that this relaxed optimization problem has an equivalent linearly constrained optimization with structured regularization that encourages approximately low rank solutions. Specializing to structures arising in neural networks, we gain insights into how the features and thus the kernel function change, providing additional nuance to the phenomenon of kernel alignment when the target function is poorly represented by tangent features. In addition to verifying our theoretical observations in real neural networks on a simple regression problem, we empirically show that an adaptive feature implementation of tangent feature classification has an order of magnitude lower sample complexity than the fixed tangent feature model on MNIST and CIFAR-10. | [
"adaptive",
"kernel learning",
"tangent kernel",
"neural networks",
"low rank"
] | https://openreview.net/pdf?id=NFqCYfoLw6 | ZTxON2cQKw | official_review | 1,696,635,326,764 | NFqCYfoLw6 | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission32/Reviewer_45h3"
] | title: Review of Submission 32: An Adaptive Tangent Feature Perspective of Neural Networks
review: ## Overview:
In this paper, the authors "propose a framework for understanding linear models in tangent feature space where the features are transformed". This enables them to gain new insight on settings that have challenged the traditional neural tangent kernel analysis, particularly in realistic, finite size neural networks that are trained on standard data (e.g., MNIST, CIFAR-10).
## Strengths:
1. The paper deals with an interesting and important goal of extending analysis beyond the NTK, where the gradients are fixed.
2. The authors develop a general framework, which enables them to study a range of problems.
3. The authors are upfront with their models' assumptions and limitations. They should be applauded for their honesty and directness.
4. The analysis provides enhanced (by an *order* of magnitude) low-sample performance of feature learning, as compared to the traditional fixed gradient analysis.
## Weaknesses:
In general, I found this paper challenging to follow. I think this is mostly a failing on my end (hence, my low confidence), but I do think there are several ways in which the authors could make their work more clear.
1. Low-rankedness: How do the assumptions shape the identified "low-rank perturbation" in Theorem 1? Is it a direct consequence of the assumptions, or a more general aspect of the adaptive feature perspective? I am also unclear whether the low-rank perturbation necessarily counts as a low-rank solution, as the phrasing in the abstract suggests ("with structured regularization that encourages approximately low rank solutions" [lines 6-7]).
2. A little more detail on how, in the structureless feature learning case (where features are able to evolve) the same solution as the standard NTK analysis emerges, would be helpful. From the motivation of the work in the introduction, I found this surprising, as I was under the assumption that changes in features would necessarily lead to a departure from the NTK.
3. More discussion surrounding the results of Theorem 2 would also be helpful. I found them challenging to interpret, and the numerical examples, while seemingly promising, were not explained in as much detail as I needed to fully understand them. Particularly, I did not feel like I fully understood Figure 3.
## Summary:
This paper tackles an interesting and important question, which is relevant and aligned CPAL's research areas. While the numerical results seemed promising, I did not feel like I fully understood them, nor the analytical theory. If the authors could provide some additional insight, I would very much consider increasing my score.
rating: 6: Marginally above acceptance threshold
confidence: 2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper |
NFqCYfoLw6 | An Adaptive Tangent Feature Perspective of Neural Networks | [
"Daniel LeJeune",
"Sina Alemohammad"
] | In order to better understand feature learning in neural networks, we propose and study linear models in tangent feature space where the features are allowed to be transformed during training. We consider linear feature transformations, resulting in a joint optimization over parameters and transformations with a bilinear interpolation constraint. We show that this relaxed optimization problem has an equivalent linearly constrained optimization with structured regularization that encourages approximately low rank solutions. Specializing to structures arising in neural networks, we gain insights into how the features and thus the kernel function change, providing additional nuance to the phenomenon of kernel alignment when the target function is poorly represented by tangent features. In addition to verifying our theoretical observations in real neural networks on a simple regression problem, we empirically show that an adaptive feature implementation of tangent feature classification has an order of magnitude lower sample complexity than the fixed tangent feature model on MNIST and CIFAR-10. | [
"adaptive",
"kernel learning",
"tangent kernel",
"neural networks",
"low rank"
] | https://openreview.net/pdf?id=NFqCYfoLw6 | S2ELybySSP | meta_review | 1,699,573,591,886 | NFqCYfoLw6 | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission32/Area_Chair_v4PM"
] | metareview: This paper proposes an extension of neural tangent kernel framework where the tangent features are allowed to adapt to the
data along with the regression coefficients. In particular, the authors propose jointly optimizing for a linear transformation of the tangent features and regression coefficients under an interpolation constraint. This is shown to be equivalent to a linearly constrained optimization problem with structured regularization that encourages approximately low rank solutions.
Reviews are mixed but skewing positive (5,6,7). Reviewers praised the work for analyzing a regime that goes beyond the NTK, and its insights into how the features and kernel functions may change when training neural networks. However, the clarity and accessibility of the paper seem to be a common concern among reviewers. The validity of the results was not called into question, but several reviewers thought more discussion surrounding the main theorems would be helpful, as well as more details for the numerical experiments. This was addressed directly by the authors in the rebuttal to the satisfaction of the reviewers.
Overall, despite some issues with clarity/presentation, the consensus is that this has paper important insights for training outside NTK regime and will be of interest to the CPAL audience.
recommendation: Accept (Poster)
confidence: 4: The area chair is confident but not absolutely certain |
NFqCYfoLw6 | An Adaptive Tangent Feature Perspective of Neural Networks | [
"Daniel LeJeune",
"Sina Alemohammad"
] | In order to better understand feature learning in neural networks, we propose and study linear models in tangent feature space where the features are allowed to be transformed during training. We consider linear feature transformations, resulting in a joint optimization over parameters and transformations with a bilinear interpolation constraint. We show that this relaxed optimization problem has an equivalent linearly constrained optimization with structured regularization that encourages approximately low rank solutions. Specializing to structures arising in neural networks, we gain insights into how the features and thus the kernel function change, providing additional nuance to the phenomenon of kernel alignment when the target function is poorly represented by tangent features. In addition to verifying our theoretical observations in real neural networks on a simple regression problem, we empirically show that an adaptive feature implementation of tangent feature classification has an order of magnitude lower sample complexity than the fixed tangent feature model on MNIST and CIFAR-10. | [
"adaptive",
"kernel learning",
"tangent kernel",
"neural networks",
"low rank"
] | https://openreview.net/pdf?id=NFqCYfoLw6 | OMVtWQEnsS | decision | 1,700,423,715,395 | NFqCYfoLw6 | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Program_Chairs"
] | decision: Accept (Oral)
comment: The paper aims to go beyond the NTK regime by allowing the tangent features to adapt to the data. They suggested that the optimization problem can be reformulated and lead to a structural regularization that promotes low-rank solutions. The reviews are somewhat mixed, where there is an appreciation for going beyond the NTK regime and yet, the paper's clarity has been called into questions to hinder understanding. The latter issues have been partially addressed during the rebuttal, and it is recommended that the authors further polish the writing, and include more discussions around the results.
The action PC chair for this paper is Yuejie Chi, who made the decision after carefully reading the paper as well as the comments by all reviewers and AC. The decision is agreed by all PC chairs.
title: Paper Decision |
NFqCYfoLw6 | An Adaptive Tangent Feature Perspective of Neural Networks | [
"Daniel LeJeune",
"Sina Alemohammad"
] | In order to better understand feature learning in neural networks, we propose and study linear models in tangent feature space where the features are allowed to be transformed during training. We consider linear feature transformations, resulting in a joint optimization over parameters and transformations with a bilinear interpolation constraint. We show that this relaxed optimization problem has an equivalent linearly constrained optimization with structured regularization that encourages approximately low rank solutions. Specializing to structures arising in neural networks, we gain insights into how the features and thus the kernel function change, providing additional nuance to the phenomenon of kernel alignment when the target function is poorly represented by tangent features. In addition to verifying our theoretical observations in real neural networks on a simple regression problem, we empirically show that an adaptive feature implementation of tangent feature classification has an order of magnitude lower sample complexity than the fixed tangent feature model on MNIST and CIFAR-10. | [
"adaptive",
"kernel learning",
"tangent kernel",
"neural networks",
"low rank"
] | https://openreview.net/pdf?id=NFqCYfoLw6 | AOWFIAdj6z | official_review | 1,696,205,632,123 | NFqCYfoLw6 | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission32/Reviewer_5sab"
] | title: Report on An Adaptive Tangent Feature Perspective of Neural Networks
review: In this paper, the authors explore the application of linear transformations to features, leading to a joint optimization encompassing parameters and transformations, while adhering to a bilinear interpolation constraint. Additionally, they demonstrate how constraining adaptivity imparts specific regularization characteristics to the solution, resulting in a group approximate low-rank penalty on a neural network-based model. This provides a framework for understanding the properties of neural networks. In summary, this paper presents an interesting and valuable contribution.
Comments:
1. In Line 146, you assume that $M$ is symmetric positive semidefinite, but it would be helpful to explain how this condition is guaranteed through the rotation of $\theta$. Furthermore, I am curious if your conclusions are applicable to symmetric matrices, which are more commonly encountered in practical implementations.
2. Reviewers have suggested that the authors provide a simplified proof sketch in main paper to assist readers in grasping the key points of the proof more easily. This addition would enhance the paper's accessibility and comprehension.
rating: 6: Marginally above acceptance threshold
confidence: 3: The reviewer is fairly confident that the evaluation is correct |
NFqCYfoLw6 | An Adaptive Tangent Feature Perspective of Neural Networks | [
"Daniel LeJeune",
"Sina Alemohammad"
] | In order to better understand feature learning in neural networks, we propose and study linear models in tangent feature space where the features are allowed to be transformed during training. We consider linear feature transformations, resulting in a joint optimization over parameters and transformations with a bilinear interpolation constraint. We show that this relaxed optimization problem has an equivalent linearly constrained optimization with structured regularization that encourages approximately low rank solutions. Specializing to structures arising in neural networks, we gain insights into how the features and thus the kernel function change, providing additional nuance to the phenomenon of kernel alignment when the target function is poorly represented by tangent features. In addition to verifying our theoretical observations in real neural networks on a simple regression problem, we empirically show that an adaptive feature implementation of tangent feature classification has an order of magnitude lower sample complexity than the fixed tangent feature model on MNIST and CIFAR-10. | [
"adaptive",
"kernel learning",
"tangent kernel",
"neural networks",
"low rank"
] | https://openreview.net/pdf?id=NFqCYfoLw6 | 6ojr3aVvSL | official_review | 1,696,657,545,331 | NFqCYfoLw6 | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission32/Reviewer_L6Tc"
] | title: An Adaptive Tangent Feature Perspective of Neural Networks
review: This paper studies linear models in tangent feature space with transformable features. This work indicates the relationship between linear feature adaptivity and structured regression using fixed features with low-rank constraints. The authors provide insights into how the features and kernel functions change. Experiments support the results. My concerns are the following.
1. This paper is not very easy to follow. The writing needs improvement.
2. There are too many assumptions about the Adaptive feature model and the neural network model. I feel it makes the conclusion weak.
3. The practical significance of the results is not very clear. For example, if the authors want to relate the results with LoRA, it is better to provide a Corollary with a discussion on LoRA.
----------------------------------------------------------------------------------------------------------------------------
Thank you for the reply, and sorry for the late reply. I am partially satisfied with the response. For the examples, I am referring to whether your assumption applies to some common neural networks, such as convolutional networks with ReLU activations. It is difficult to make a decision. I will keep my rate of 5 but decrease my confidence to 2.
rating: 5
confidence: 2 |
MxBS6aw5Gd | Probing Biological and Artificial Neural Networks with Task-dependent Neural Manifolds | [
"Michael Kuoch",
"Chi-Ning Chou",
"Nikhil Parthasarathy",
"Joel Dapello",
"James J. DiCarlo",
"Haim Sompolinsky",
"SueYeon Chung"
] | In recent years, growth in our understanding of the computations performed in both biological and artificial neural networks has largely been driven by either low-level mechanistic studies or global normative approaches. However, concrete methodologies for bridging the gap between these levels of abstraction remain elusive. In this work, we investigate the internal mechanisms of neural networks through the lens of neural population geometry, aiming to provide understanding at an intermediate level of abstraction, as a way to bridge that gap. Utilizing manifold capacity theory (MCT) from statistical physics and manifold alignment analysis (MAA) from high-dimensional statistics, we probe the underlying organization of task-dependent manifolds in deep neural networks and neural recordings from the macaque visual cortex. Specifically, we quantitatively characterize how different learning objectives lead to differences in the organizational strategies of these models and demonstrate how these geometric analyses are connected to the decodability of task-relevant information. Furthermore, these metrics show that macaque visual cortex data are more similar to unsupervised DNNs in terms of geometrical properties such as manifold position and manifold alignment. These analyses present a strong direction for bridging mechanistic and normative theories in neural networks through neural population geometry, potentially opening up many future research avenues in both machine learning and neuroscience. | [
"Computational Neuroscience",
"Neural Manifolds",
"Neural Geometry",
"Representational Geometry",
"Biologically inspired vision models",
"Neuro-AI"
] | https://openreview.net/pdf?id=MxBS6aw5Gd | dn2YVHIGQ4 | decision | 1,700,497,731,819 | MxBS6aw5Gd | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Program_Chairs"
] | decision: Accept (Oral)
title: Paper Decision |
MxBS6aw5Gd | Probing Biological and Artificial Neural Networks with Task-dependent Neural Manifolds | [
"Michael Kuoch",
"Chi-Ning Chou",
"Nikhil Parthasarathy",
"Joel Dapello",
"James J. DiCarlo",
"Haim Sompolinsky",
"SueYeon Chung"
] | In recent years, growth in our understanding of the computations performed in both biological and artificial neural networks has largely been driven by either low-level mechanistic studies or global normative approaches. However, concrete methodologies for bridging the gap between these levels of abstraction remain elusive. In this work, we investigate the internal mechanisms of neural networks through the lens of neural population geometry, aiming to provide understanding at an intermediate level of abstraction, as a way to bridge that gap. Utilizing manifold capacity theory (MCT) from statistical physics and manifold alignment analysis (MAA) from high-dimensional statistics, we probe the underlying organization of task-dependent manifolds in deep neural networks and neural recordings from the macaque visual cortex. Specifically, we quantitatively characterize how different learning objectives lead to differences in the organizational strategies of these models and demonstrate how these geometric analyses are connected to the decodability of task-relevant information. Furthermore, these metrics show that macaque visual cortex data are more similar to unsupervised DNNs in terms of geometrical properties such as manifold position and manifold alignment. These analyses present a strong direction for bridging mechanistic and normative theories in neural networks through neural population geometry, potentially opening up many future research avenues in both machine learning and neuroscience. | [
"Computational Neuroscience",
"Neural Manifolds",
"Neural Geometry",
"Representational Geometry",
"Biologically inspired vision models",
"Neuro-AI"
] | https://openreview.net/pdf?id=MxBS6aw5Gd | 7MkCv27W4Y | meta_review | 1,700,017,479,272 | MxBS6aw5Gd | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission53/Area_Chair_7xU4"
] | metareview: (this is an invited paper, so only one meta-review is provided)
Machine learning, and deep learning in particular, have drawn significant
inspiration from neuroscience over the years, with far-reaching impacts.
This submission, along with an emerging body of exciting research, demonstrates
that this statement can also be turned around: deep learning can
provide insights and inspirations for research into the nature of representation
learning in both artificial and biological neural networks.
To this end, the authors conduct a detailed study into how representations in
artificial (ResNet backbone) and biological (in vivo micro-electrode recordings
of V4 and IT in macaques) neural networks differ from one another as a function
of task-specific aspects, such as supervised versus unsupervised training as
well as object-specific or class-specific conditioning.
They base this study on computational metrics related to the geometry of
representations -- specifically, metrics from manifold capacity theory (MCT),
advanced by Chung et al. (2018) with tools from statistical physics, are used to
quantify how well representations of different classes can be separated from one
another, and metrics from manifold alignment analysis (MAA) are used/proposed to
quantify how well representations of different classes align with one another
(e.g., in the sense of disentangled representation learning). The work
contributes to MAA by proposing a new metric, the signal mismatch distance, with
detailed calculations and verifications in simple models given in the
supplemental.
Their experiments obtain artificial and biological neural representations via
stimulation using visual inputs consisting of one of a set of controlled 3D
objects (6D pose) embedded into a background, for the different pretrained deep
networks and macaques.
Within this context, the authors present three classes of findings.
The first is that unsupervised learning methods for deep networks (BYOL, SimCLR,
many more)
learn representations with systematically lower MCT scores than supervised
learning methods, suggesting less task-specialized (i.e., classification)
behavior of the learned representations; these findings correlate with the
observations for V4 (earlier in the visual cortex) and IT. The second is that
unsupervised deep network representations display greater MAA scores than
supervised representations, and that this extends to unsupervised deep
representations showing greater alignment similarity to neural representations
from the macaque visual cortex. Interestingly, these experiments demonstrate
that a standard neural data analysis technique fails to distinguish these
similarities -- it is necessary to leverage task-specific information in the
comparison between artificial and biological representations, as in the MAA
metrics. The third is that when controlling the visual stimulus to vary along a
single degree of freedom (fixed object, varying pose), representational
alignment (measured through MAA metrics) typically correlates positively with
various indicators of downstream performance.
Taken together, these results suggest that metrics based on data geometry
provide a useful avenue to analyze representations in biological and artificial
neural networks, as well as interesting avenues for future work, especially
surrounding better understanding of the novel MAA metrics.
These results provide a fresh perspective on representation learning, amidst
other emerging theories for representation learning based on, for example,
neural collapse and compression, and will make a valuable addition to CPAL.
For the camera-ready, the authors may wish to correct a few minor points:
removing lines 465-466; changing $\\| AB \\|$
to $\\| A^{1/2} B^{1/2} \\|$ (both nuclear norms) in eqn. (4).
recommendation: Accept (Oral)
confidence: 5: The area chair is absolutely certain |
MlgnGWdqWl | Exploring Minimally Sufficient Representation in Active Learning through Label-Irrelevant Patch Augmentation | [
"Zhiyu Xue",
"Yinlong Dai",
"Qi Lei"
] | Deep learning models, which require abundant labeled data for training, are expensive and time-consuming to implement, particularly in medical imaging. Active learning (AL) aims to maximize model performance with few labeled samples by gradually expanding and labeling a new training set. In this work, we intend to learn a "good" feature representation that is both sufficient and minimal, facilitating effective AL for medical image classification. This work proposes an efficient AL framework based on off-the-shelf self-supervised learning models, complemented by a label-irrelevant patch augmentation scheme. This scheme is designed to reduce redundancy in the learned features and mitigate overfitting in the progress of AL. Our framework offers efficiency to AL in terms of parameters, samples, and computational costs. The benefits of this approach are extensively validated across various medical image classification tasks employing different AL strategies.~\footnote{Source Codes: \url{https://github.com/chrisyxue/DA4AL}}. | [
"Active Learning",
"Data Augmentation",
"Minimally Sufficient Representation"
] | https://openreview.net/pdf?id=MlgnGWdqWl | mJ1q0XbcG2 | decision | 1,700,437,847,122 | MlgnGWdqWl | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Program_Chairs"
] | decision: Accept (Oral)
comment: After a careful reconsideration and taking into account the reviewers' comments and concerns, we have decided to accept the paper to the conference.
The paper addresses a significant challenge in the field of medical image classification, namely, the need for ample labeled data for training deep learning models. The focus on active learning (AL) and the development of a feature representation that balances sufficiency and minimality is commendable.
The authors have proposed an efficient AL framework that leverages self-supervised learning models and introduced a patch augmentation scheme to enhance the AL process. While there were concerns raised by the reviewers, the authors' responses have addressed some of these concerns.
Upon further consideration, I believe that the paper presents novel and interesting formulations and results, even though some concerns remain. Given the potential value of the proposed approach and the authors' efforts to address the reviewers' comments, we recommend the paper for acceptance. However, we encourage the authors to continue refining their work and addressing the remaining concerns to further improve the paper's quality before the camera-ready.
The action PC chair for this paper is Qing Qu, who made the decision after carefully reading the paper as well as the comments by all reviewers and AC. The decision is agreed upon by all PC chairs.
title: Paper Decision |
MlgnGWdqWl | Exploring Minimally Sufficient Representation in Active Learning through Label-Irrelevant Patch Augmentation | [
"Zhiyu Xue",
"Yinlong Dai",
"Qi Lei"
] | Deep learning models, which require abundant labeled data for training, are expensive and time-consuming to implement, particularly in medical imaging. Active learning (AL) aims to maximize model performance with few labeled samples by gradually expanding and labeling a new training set. In this work, we intend to learn a "good" feature representation that is both sufficient and minimal, facilitating effective AL for medical image classification. This work proposes an efficient AL framework based on off-the-shelf self-supervised learning models, complemented by a label-irrelevant patch augmentation scheme. This scheme is designed to reduce redundancy in the learned features and mitigate overfitting in the progress of AL. Our framework offers efficiency to AL in terms of parameters, samples, and computational costs. The benefits of this approach are extensively validated across various medical image classification tasks employing different AL strategies.~\footnote{Source Codes: \url{https://github.com/chrisyxue/DA4AL}}. | [
"Active Learning",
"Data Augmentation",
"Minimally Sufficient Representation"
] | https://openreview.net/pdf?id=MlgnGWdqWl | kTjSgrOWi5 | meta_review | 1,699,846,191,143 | MlgnGWdqWl | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission8/Area_Chair_Lhk8"
] | metareview: In this paper, the authors address the challenge of obtaining ample labeled data for training deep learning models, particularly in the context of medical imaging, where data collection is costly and time-consuming. They focus on the concept of active learning (AL), which aims to maximize model performance using a limited number of labeled samples by iteratively expanding and labeling the training dataset. The primary objective of this study is to develop a feature representation that strikes a balance between sufficiency and minimality, thereby facilitating effective AL for medical image classification.
The authors propose an efficient AL framework that leverages readily available self-supervised learning models. Additionally, they introduce a patch augmentation scheme that they claim is insensitive to labels, aiming to reduce redundancy in the learned features and mitigate overfitting during the AL process. The authors also claim to validate the performance improvements achieved by their approach across various medical image classification tasks using different AL strategies. The reviewers thought some of the formulation and results were novel and interesting but raised various concerns including expense of training an adaptor, marginal benefits, limited experiments (even in the medical domain), lack of intuition, lack of precise comparisons. While the authors response addressed some of these concerns it is clear that many remain. All reviewers agree that the paper is marginally below the threshold and I concur and therefore cannot recommend acceptance at this time.
recommendation: Reject
confidence: 5: The area chair is absolutely certain |
MlgnGWdqWl | Exploring Minimally Sufficient Representation in Active Learning through Label-Irrelevant Patch Augmentation | [
"Zhiyu Xue",
"Yinlong Dai",
"Qi Lei"
] | Deep learning models, which require abundant labeled data for training, are expensive and time-consuming to implement, particularly in medical imaging. Active learning (AL) aims to maximize model performance with few labeled samples by gradually expanding and labeling a new training set. In this work, we intend to learn a "good" feature representation that is both sufficient and minimal, facilitating effective AL for medical image classification. This work proposes an efficient AL framework based on off-the-shelf self-supervised learning models, complemented by a label-irrelevant patch augmentation scheme. This scheme is designed to reduce redundancy in the learned features and mitigate overfitting in the progress of AL. Our framework offers efficiency to AL in terms of parameters, samples, and computational costs. The benefits of this approach are extensively validated across various medical image classification tasks employing different AL strategies.~\footnote{Source Codes: \url{https://github.com/chrisyxue/DA4AL}}. | [
"Active Learning",
"Data Augmentation",
"Minimally Sufficient Representation"
] | https://openreview.net/pdf?id=MlgnGWdqWl | eQq5pFxi1i | official_review | 1,696,697,048,207 | MlgnGWdqWl | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission8/Reviewer_VkB2"
] | title: Official Review of Submission 8
review: # Overview
The manuscript proposed a new framework to improve the performance of active learning, specifically for medical image classification.
The proposed method is based on a self-supervised model and a label-irrelevant patch augmentation scheme. Plenty of ablation studies have been conducted, showing the effectiveness of the proposed method.
# Comments
(+) Thorough ablation studies are carried out, providing a deep understanding of the method's effectiveness.
(+) The method is articulated clearly and supported with descriptive diagrams, enhancing comprehension.
(-) Regarding the results presented in Table 2, it's unclear if the FLOPs consider the overall costs or only account for individual rounds. Given that every AL round necessitates training a new adapter based on the current encoder, the overall computational overhead seems not negligible.
(-) The abstract mentioned, "to reduce redundancy in the learned features and mitigate overfitting during the AL process". Are there specific results that demonstrate the successful mitigation of overfitting?
(-) The advancements illustrated in Table 1 appear to be relatively mild. Incorporating error bars might offer a more effective assessment of the proposed method.
rating: 5: Marginally below acceptance threshold
confidence: 3: The reviewer is fairly confident that the evaluation is correct |
MlgnGWdqWl | Exploring Minimally Sufficient Representation in Active Learning through Label-Irrelevant Patch Augmentation | [
"Zhiyu Xue",
"Yinlong Dai",
"Qi Lei"
] | Deep learning models, which require abundant labeled data for training, are expensive and time-consuming to implement, particularly in medical imaging. Active learning (AL) aims to maximize model performance with few labeled samples by gradually expanding and labeling a new training set. In this work, we intend to learn a "good" feature representation that is both sufficient and minimal, facilitating effective AL for medical image classification. This work proposes an efficient AL framework based on off-the-shelf self-supervised learning models, complemented by a label-irrelevant patch augmentation scheme. This scheme is designed to reduce redundancy in the learned features and mitigate overfitting in the progress of AL. Our framework offers efficiency to AL in terms of parameters, samples, and computational costs. The benefits of this approach are extensively validated across various medical image classification tasks employing different AL strategies.~\footnote{Source Codes: \url{https://github.com/chrisyxue/DA4AL}}. | [
"Active Learning",
"Data Augmentation",
"Minimally Sufficient Representation"
] | https://openreview.net/pdf?id=MlgnGWdqWl | Rz02WRSDsG | official_review | 1,696,728,163,413 | MlgnGWdqWl | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission8/Reviewer_2XdG"
] | title: Review for Exploring Minimally Sufficient Representation in Active Learning through Label-Irrelevant Patch Augmentation
review: Summary Of The Paper
This paper studies active learning (AL) for medical image classification. The authors introduce an efficient AL framework based on self-supervised learning models. This paper also introduces a patch augmentation method to enhance the feature representation quality and reduce redundancy, ultimately preventing overfitting during AL.
Main Review
- Strength
1) The paper is written in a clear manner. The authors provide a comprehensive background and give clear justification and problem setting for their proposed work.
2) In addition, the proposed method is supported by multiple datasets, and the authors have conducted extensive ablation studies, including several versions of augmentations, which is good to justify the proposed model.
- Weakness
1) With extensive experiments and analysis, the paper has demonstrated its proposed strategy on several dataset, however, the intuition of the proposed method is not thoroughly discussed. I recommend the authors should elaborate more on the design of their label-irrelevant patches augmentations.
2) In order to demonstrate the effectiveness of the proposed method, downstream tasks such as detection on AL (e.g., https://github.com/yuantn/MI-AOD) could be explored. I think the proposed method may not be practical without other tasks.
3) Have the authors investigated and addressed potential domain discrepancies between the pretrained ViTs, which are trained on natural images, and their fine-tuned adaptor with medical images (specific domain)?
4) I'm not entirely certain if Table 2 provides an apples-to-apples comparison. In one case, the entire VIT backbone is frozen, and only the adapter is fine-tuned. However, in the other case, Resnet18-50 is trained end-to-end. What would happen if I were to take a pretrained Resnet18 and then add an adapter for fine-tuning?
5) I suggest the authors to provide a comprehensive survey on relevant works such as augmentation strategy on representation learning (including self-supervised learning), For example: [A] Kügelgen et al, Self-Supervised Learning with Data Augmentations Provably Isolates Content from Style, [B] Yeh et. al., SAGA: Self-Augmentation with Guided Attention for Representation Learning
Summary Of The Review
Overall, without discussing the intuition of the proposed method in details, I am not sure if the novelty that authors mention in the paper is reliable. Also, I think we need to see valid elaboration and the intuition of the proposed method, and scale up to other tasks. Combined with the weaknesses I mentioned above, I vote for 5. I would like to see authors response to consider raising the rating.
rating: 5: Marginally below acceptance threshold
confidence: 4: The reviewer is confident but not absolutely certain that the evaluation is correct |
MlgnGWdqWl | Exploring Minimally Sufficient Representation in Active Learning through Label-Irrelevant Patch Augmentation | [
"Zhiyu Xue",
"Yinlong Dai",
"Qi Lei"
] | Deep learning models, which require abundant labeled data for training, are expensive and time-consuming to implement, particularly in medical imaging. Active learning (AL) aims to maximize model performance with few labeled samples by gradually expanding and labeling a new training set. In this work, we intend to learn a "good" feature representation that is both sufficient and minimal, facilitating effective AL for medical image classification. This work proposes an efficient AL framework based on off-the-shelf self-supervised learning models, complemented by a label-irrelevant patch augmentation scheme. This scheme is designed to reduce redundancy in the learned features and mitigate overfitting in the progress of AL. Our framework offers efficiency to AL in terms of parameters, samples, and computational costs. The benefits of this approach are extensively validated across various medical image classification tasks employing different AL strategies.~\footnote{Source Codes: \url{https://github.com/chrisyxue/DA4AL}}. | [
"Active Learning",
"Data Augmentation",
"Minimally Sufficient Representation"
] | https://openreview.net/pdf?id=MlgnGWdqWl | 3I3YQOAiN0 | official_review | 1,697,402,005,350 | MlgnGWdqWl | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission8/Reviewer_XeA1"
] | title: Review for paper 8
review: The paper proposes an efficient AL framework based on off-the-shelf self-supervised learning models, and a label-irrelevant patch augmentation scheme is introduced to reduce redundancy in the learned features and mitigate overfitting in the progress of Active Learning. Some key positives and weaknesses are listed as:
Strength:
1. Their formulation of a label-irrelevant patch augmentation scheme that preserves semantic information is interesting.
2. The proposed parameter-efficient AL framework can boost the overall performance of Few-shot AL by 5% − 7%.
3. The paper is very well written and it includes many interesting ablation studies + experimental details (eg. aug analysis in the appendix).
Weakness:
1. I am concerned about the expense of training an adapter on top of a pre-trained and fixed ViT in every AL round. In what way, this additional training cost can compensate for the performance gain?
2. Performance benefits of the proposed approach seem very marginal and sometimes less than RandAug. Can authors provide some explanation?
3. Additional results on non-medical images on conventional CV datasets can provide more evidence of benefits. I am not sure why authors limited their evaluation to medical images. Some widely popular medical datasets like ChestX-rays etc. are missing.
rating: 5: Marginally below acceptance threshold
confidence: 4: The reviewer is confident but not absolutely certain that the evaluation is correct |
MF35ZwzpY5 | Unsupervised Learning of Structured Representation via Closed-Loop Transcription | [
"Shengbang Tong",
"Xili Dai",
"Yubei Chen",
"Mingyang Li",
"ZENGYI LI",
"Brent Yi",
"Yann LeCun",
"Yi Ma"
] | This paper proposes an unsupervised method for learning a unified representation that serves both discriminative and generative purposes. While most existing unsupervised learning approaches focus on a representation for only one of these two goals, we show that a unified representation can enjoy the mutual benefits of having both. Such a representation is attainable by generalizing the recently proposed closed-loop transcription framework, known as CTRL, to the unsupervised setting. This entails solving a constrained maximin game over a rate reduction objective that expands features of all samples while compressing features of augmentations of each sample. Through this process, we see discriminative low-dimensional structures emerge in the resulting representations. Under comparable experimental conditions and network complexities, we demonstrate that these structured representations enable classification performance close to state-of-the-art unsupervised discriminative representations, and conditionally generated image quality significantly higher than that of state-of-the-art unsupervised generative models. | [
"Unsupervised/Self-supervised Learning",
"Closed-Loop Transcription"
] | https://openreview.net/pdf?id=MF35ZwzpY5 | lPEKi1Tv5O | official_review | 1,697,403,033,029 | MF35ZwzpY5 | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission6/Reviewer_B6YQ"
] | title: Interesting work about an unsupervised learning method for discriminative and generative tasks
review: This paper proposes an unsupervised representation learning method for generative and discriminative tasks.
(1) The quality is high, given abundant experiments and visualization on classification accuracy and representation learning.
(2) The clarity is good. The paper is overall easy to follow.
(3) The originality is fair since the proposed methods, such as the CTRL-binary program and self-supervised learning with data augmentation, are extensions of existing works, although the authors properly cite them.
(4) The significance is good, and I think this work attracts the community.
rating: 6: Marginally above acceptance threshold
confidence: 2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper |
MF35ZwzpY5 | Unsupervised Learning of Structured Representation via Closed-Loop Transcription | [
"Shengbang Tong",
"Xili Dai",
"Yubei Chen",
"Mingyang Li",
"ZENGYI LI",
"Brent Yi",
"Yann LeCun",
"Yi Ma"
] | This paper proposes an unsupervised method for learning a unified representation that serves both discriminative and generative purposes. While most existing unsupervised learning approaches focus on a representation for only one of these two goals, we show that a unified representation can enjoy the mutual benefits of having both. Such a representation is attainable by generalizing the recently proposed closed-loop transcription framework, known as CTRL, to the unsupervised setting. This entails solving a constrained maximin game over a rate reduction objective that expands features of all samples while compressing features of augmentations of each sample. Through this process, we see discriminative low-dimensional structures emerge in the resulting representations. Under comparable experimental conditions and network complexities, we demonstrate that these structured representations enable classification performance close to state-of-the-art unsupervised discriminative representations, and conditionally generated image quality significantly higher than that of state-of-the-art unsupervised generative models. | [
"Unsupervised/Self-supervised Learning",
"Closed-Loop Transcription"
] | https://openreview.net/pdf?id=MF35ZwzpY5 | UDrwCsMCI9 | official_review | 1,697,411,783,554 | MF35ZwzpY5 | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission6/Reviewer_Vsca"
] | title: Review
review: **Summary:**
The paper introduces a method for learning representations that are applicable to both discriminative and generative tasks, employing a closed-loop transaction framework. Building on the groundwork laid by [2] (referenced as [16] in this paper), which proposes a rate reduction objective to quantify the disparity between encoded representations, this paper identifies a limitation in this approach - the absence of sample-wise self-consistency. This concern is addressed through the introduction of a specialized rate reduction loss. Additionally, the authors employ another loss term to adapt the CTRL framework for unsupervised settings.
**Pros:**
- The proposed method demonstrates versatility, being applicable to both discriminative tasks and image generation, including unsupervised conditional image generation.
- The implementation of self-supervision, requiring that the sample and its augmentations are closely situated in the encoded space as measured by the rate reduction, is a straightforward yet interesting concept.
**Cons:**
- The proposed modifications appear to be largely incremental in comparison to the work in [2], primarily revolving around the incorporation of specialized losses to influence the encoded space. Notably, such losses have previously been discussed in the literature (see [1] in the References below for self-consistency, and „Detailed” notes).
- The experiments primarily focus on the CIFAR-10/100 and Tiny-ImageNet datasets, which I perceive as a somewhat limited evaluation. This is particularly notable given that CIFAR-100 shares a largely similar semantic structure with CIFAR-10. It would be preferable to see a more extensive evaluation on additional datasets. Specifically, it would be valuable to assess whether the method scales effectively for larger datasets or images.
**Detailed:**
My main concern with this work is that it appears to represent a modest incremental advancement over existing approaches, particularly [1] and [2]. Both of these works consider rate reduction for representation learning, with [1] even placing particular emphasis on the self-consistency of the representation. Consequently, I find the contribution of this paper to be somewhat constrained.
**References:**
[1] Ma, Yi, Doris Tsao, and Heung-Yeung Shum. "On the principles of parsimony and self-consistency for the emergence of intelligence." Frontiers of Information Technology & Electronic Engineering 23.9 (2022): 1298-1323.
[2] Dai, X., Tong, S., Li, M., Wu, Z., Psenka, M., Chan, K. H. R., ... & Ma, Y. (2022). Ctrl: Closed-loop transcription to an ldr via minimaxing rate reduction. Entropy, 24(4), 456.
rating: 5: Marginally below acceptance threshold
confidence: 2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper |
MF35ZwzpY5 | Unsupervised Learning of Structured Representation via Closed-Loop Transcription | [
"Shengbang Tong",
"Xili Dai",
"Yubei Chen",
"Mingyang Li",
"ZENGYI LI",
"Brent Yi",
"Yann LeCun",
"Yi Ma"
] | This paper proposes an unsupervised method for learning a unified representation that serves both discriminative and generative purposes. While most existing unsupervised learning approaches focus on a representation for only one of these two goals, we show that a unified representation can enjoy the mutual benefits of having both. Such a representation is attainable by generalizing the recently proposed closed-loop transcription framework, known as CTRL, to the unsupervised setting. This entails solving a constrained maximin game over a rate reduction objective that expands features of all samples while compressing features of augmentations of each sample. Through this process, we see discriminative low-dimensional structures emerge in the resulting representations. Under comparable experimental conditions and network complexities, we demonstrate that these structured representations enable classification performance close to state-of-the-art unsupervised discriminative representations, and conditionally generated image quality significantly higher than that of state-of-the-art unsupervised generative models. | [
"Unsupervised/Self-supervised Learning",
"Closed-Loop Transcription"
] | https://openreview.net/pdf?id=MF35ZwzpY5 | SSTjptBFpP | official_review | 1,697,402,064,624 | MF35ZwzpY5 | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission6/Reviewer_E7S2"
] | title: Review for paper 6
review: The paper proposes an unsupervised method for learning a unified representation that serves both discriminative and generative purposes enjoying the mutual benefits of having both. Their work illustrates that the structured representation learned is discriminative and simple classifiers applied to learned features yield high classification accuracy. In addition, it poses enough diversity to recover raw inputs, and
46 structure that can be exploited for sampling and generating new images. Some key positive and weakness are listed as:
Strength:
1. Their formulation benefits from the mutual benefits of both discriminative and generative properties.
2. It is interesting that closed-loop transcription through the maximin game between the encoder and decoder has the potential to offer a unifying framework for both discriminative and generative representation learning, across supervised, incremental, and unsupervised settings.
3. The paper is very well written and it includes many interesting ablation studies + experimental details (eg. cluster analysis in the appendix).
Weakness:
1. Table3 indicate that the performance U-CTRL still lags behind in comparison with traditional purely discriminative self-supervised learning methods like SimCLR, MoCo, etc. Although, I agree that it carry the benefits of both world. Can authors add more recent beaslines like oCoV3 etc.
2. Results should be included with standard deviation, and how closely the evaluation setting of baselines relate with the proposed method is not yet clear.
3. Additional results on classical image generation datasets will make the paper more convincing.
rating: 7: Good paper, accept
confidence: 3: The reviewer is fairly confident that the evaluation is correct |
MF35ZwzpY5 | Unsupervised Learning of Structured Representation via Closed-Loop Transcription | [
"Shengbang Tong",
"Xili Dai",
"Yubei Chen",
"Mingyang Li",
"ZENGYI LI",
"Brent Yi",
"Yann LeCun",
"Yi Ma"
] | This paper proposes an unsupervised method for learning a unified representation that serves both discriminative and generative purposes. While most existing unsupervised learning approaches focus on a representation for only one of these two goals, we show that a unified representation can enjoy the mutual benefits of having both. Such a representation is attainable by generalizing the recently proposed closed-loop transcription framework, known as CTRL, to the unsupervised setting. This entails solving a constrained maximin game over a rate reduction objective that expands features of all samples while compressing features of augmentations of each sample. Through this process, we see discriminative low-dimensional structures emerge in the resulting representations. Under comparable experimental conditions and network complexities, we demonstrate that these structured representations enable classification performance close to state-of-the-art unsupervised discriminative representations, and conditionally generated image quality significantly higher than that of state-of-the-art unsupervised generative models. | [
"Unsupervised/Self-supervised Learning",
"Closed-Loop Transcription"
] | https://openreview.net/pdf?id=MF35ZwzpY5 | PNMt0h41Lz | decision | 1,700,363,047,324 | MF35ZwzpY5 | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Program_Chairs"
] | decision: Accept (Oral)
comment: The paper introduces the Unsupervised-CTRL (U-CTRL) framework for unsupervised representation learning applicable to both generative and discriminative tasks. Reviewers appreciate the extensive experiments, clarity, and versatility of the proposed method. However, some concerns include the perception that the contributions may be somewhat incremental compared to existing approaches and the limited evaluation on benchmark datasets. The authors are encouraged to explore its real-world applicability on diverse datasets and further optimize hyperparameters for U-CTRL. Overall, the paper is seen as high-quality and significant.
The action PC chair for this paper is Atlas Wang, who made the decision after carefully reading the paper as well as the comments by all reviewers and AC. The decision is agreed by all PC chairs.
title: Paper Decision |
MF35ZwzpY5 | Unsupervised Learning of Structured Representation via Closed-Loop Transcription | [
"Shengbang Tong",
"Xili Dai",
"Yubei Chen",
"Mingyang Li",
"ZENGYI LI",
"Brent Yi",
"Yann LeCun",
"Yi Ma"
] | This paper proposes an unsupervised method for learning a unified representation that serves both discriminative and generative purposes. While most existing unsupervised learning approaches focus on a representation for only one of these two goals, we show that a unified representation can enjoy the mutual benefits of having both. Such a representation is attainable by generalizing the recently proposed closed-loop transcription framework, known as CTRL, to the unsupervised setting. This entails solving a constrained maximin game over a rate reduction objective that expands features of all samples while compressing features of augmentations of each sample. Through this process, we see discriminative low-dimensional structures emerge in the resulting representations. Under comparable experimental conditions and network complexities, we demonstrate that these structured representations enable classification performance close to state-of-the-art unsupervised discriminative representations, and conditionally generated image quality significantly higher than that of state-of-the-art unsupervised generative models. | [
"Unsupervised/Self-supervised Learning",
"Closed-Loop Transcription"
] | https://openreview.net/pdf?id=MF35ZwzpY5 | HFZp8WOjRE | meta_review | 1,699,831,080,393 | MF35ZwzpY5 | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission6/Area_Chair_GPcZ"
] | metareview: The authors did a great job addressing the reviewers' concerns, and all reviewers but one agreed that the paper was above the acceptance threshold.
The only comment by the reviewers that remains to be addressed by the authors is how their work differs from the one of [1].
[1] Ma, Yi, Doris Tsao, and Heung-Yeung Shum. "On the principles of parsimony and self-consistency for the emergence of intelligence." Frontiers of Information Technology & Electronic Engineering 23.9 (2022): 1298-1323.
recommendation: Accept (Poster)
confidence: 3: The area chair is somewhat confident |
MF35ZwzpY5 | Unsupervised Learning of Structured Representation via Closed-Loop Transcription | [
"Shengbang Tong",
"Xili Dai",
"Yubei Chen",
"Mingyang Li",
"ZENGYI LI",
"Brent Yi",
"Yann LeCun",
"Yi Ma"
] | This paper proposes an unsupervised method for learning a unified representation that serves both discriminative and generative purposes. While most existing unsupervised learning approaches focus on a representation for only one of these two goals, we show that a unified representation can enjoy the mutual benefits of having both. Such a representation is attainable by generalizing the recently proposed closed-loop transcription framework, known as CTRL, to the unsupervised setting. This entails solving a constrained maximin game over a rate reduction objective that expands features of all samples while compressing features of augmentations of each sample. Through this process, we see discriminative low-dimensional structures emerge in the resulting representations. Under comparable experimental conditions and network complexities, we demonstrate that these structured representations enable classification performance close to state-of-the-art unsupervised discriminative representations, and conditionally generated image quality significantly higher than that of state-of-the-art unsupervised generative models. | [
"Unsupervised/Self-supervised Learning",
"Closed-Loop Transcription"
] | https://openreview.net/pdf?id=MF35ZwzpY5 | AfdsaD0RQQ | official_review | 1,696,715,975,389 | MF35ZwzpY5 | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission6/Reviewer_Enfn"
] | title: see below
review: Summary:
This presents the Unsupervised-CTRL (U-CTRL) framework, an innovative approach to unsupervised learning. The document thoroughly explains U-CTRL's formulation and objectives, demonstrating its ability to learn structured representations from unlabeled data. Extensive experiments on datasets like CIFAR-10 and CIFAR-100 showcase U-CTRL's effectiveness in clustering, image generation, and feature extraction tasks. Ablation studies emphasize the importance of different framework components. The document concludes by highlighting U-CTRL's robustness and contributions to unsupervised learning, making it a valuable resource for researchers and practitioners in this field.
Pros:
1. Unified Structured Representations: The framework focuses on learning structured representations, which can lead to more interpretable and meaningful features compared to traditional unsupervised methods. Moreover, compared with supervised-CTRL, this work simply view each sample as a new classes, providing a unified framework for learning structured and more interpretable features.
2. Versatile Applications: The document showcases various applications of U-CTRL, including image clustering, conditional image generation, and representation learning, highlighting its versatility across different domains and tasks.
3. Ablation Studies: The ablation studies provide valuable insights into the importance of different components within the U-CTRL framework, aiding researchers in understanding its inner workings and potential for improvement.
4. Competitive Performance: The framework achieves competitive performance on benchmark datasets like CIFAR-10 and CIFAR-100, indicating its effectiveness in comparison to other methods.
Cons:
1. Limited Real-World Data: The document mainly focuses on experiments with benchmark datasets like CIFAR-10 and CIFAR-100. The real-world applicability of U-CTRL may require further validation on diverse and complex data sources.
2. Hyperparameter Tuning: U-CTRL introduces two extra hyper-parameters $\lambda_1$ and $\lambda_2$. Finding optimal hyperparameters may require extensive experimentation.
rating: 8: Top 50% of accepted papers, clear accept
confidence: 4: The reviewer is confident but not absolutely certain that the evaluation is correct |
LzhpSfqSXC | Algorithm Design for Online Meta-Learning with Task Boundary Detection | [
"Daouda Sow",
"Sen Lin",
"Yingbin Liang",
"Junshan Zhang"
] | Online meta-learning has recently emerged as a marriage between batch meta-learning and online learning, for achieving the capability of quick adaptation on new tasks in a lifelong manner. However, most existing approaches focus on the restrictive setting where the distribution of the online tasks remains fixed with known task boundaries. In this work, we relax these assumptions and propose a novel algorithm for task-agnostic online meta-learning in non-stationary environments. More specifically, we first propose two simple but effective detection mechanisms of task switches and distribution shift based on empirical observations, which serve as a key building block for more elegant online model updates in our algorithm: the task switch detection mechanism allows reusing of the best model available for the current task at hand, and the distribution shift detection mechanism differentiates the meta model update in order to preserve the knowledge for in-distribution tasks and quickly learn the new knowledge for out-of-distribution tasks. In particular, our online meta model updates are based only on the current data, which eliminates the need of storing previous data as required in most existing methods. We further show that a sublinear task-averaged regret can be achieved for our algorithm under mild conditions. Empirical studies on three different benchmarks clearly demonstrate the significant advantage of our algorithm over related baseline approaches. | [
"online meta-learning",
"task boundary detection",
"domain shift",
"dynamic regret",
"out of distribution detection"
] | https://openreview.net/pdf?id=LzhpSfqSXC | afcb1TmVR0 | meta_review | 1,699,840,345,588 | LzhpSfqSXC | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission51/Area_Chair_vZNt"
] | metareview: (This is an invited paper, so only one meta-review is provided)
This work proposed a new, task-agnostic online meta-learning algorithm for dynamic environments. It addresses key challenges in the field by allowing for the algorithm to operate without fixed task distributions or known task boundaries. The introduction of mechanisms to detect task switches and distribution shifts based on empirical observations is particularly commendable. These mechanisms are crucial for the algorithm's adaptability, enabling it to maintain performance on in-distribution tasks while swiftly adapting to new, out-of-distribution tasks. The avoidance of data storage from previous tasks and the demonstration of sublinear task-averaged regret are also impressive features that suggest significant improvements over existing approaches. Empirical validation across multiple benchmarks substantiates the algorithm's efficacy. Therefore, the paper appears to be a valuable contribution and is recommended for acceptance.
recommendation: Accept (Oral)
confidence: 4: The area chair is confident but not absolutely certain |
LzhpSfqSXC | Algorithm Design for Online Meta-Learning with Task Boundary Detection | [
"Daouda Sow",
"Sen Lin",
"Yingbin Liang",
"Junshan Zhang"
] | Online meta-learning has recently emerged as a marriage between batch meta-learning and online learning, for achieving the capability of quick adaptation on new tasks in a lifelong manner. However, most existing approaches focus on the restrictive setting where the distribution of the online tasks remains fixed with known task boundaries. In this work, we relax these assumptions and propose a novel algorithm for task-agnostic online meta-learning in non-stationary environments. More specifically, we first propose two simple but effective detection mechanisms of task switches and distribution shift based on empirical observations, which serve as a key building block for more elegant online model updates in our algorithm: the task switch detection mechanism allows reusing of the best model available for the current task at hand, and the distribution shift detection mechanism differentiates the meta model update in order to preserve the knowledge for in-distribution tasks and quickly learn the new knowledge for out-of-distribution tasks. In particular, our online meta model updates are based only on the current data, which eliminates the need of storing previous data as required in most existing methods. We further show that a sublinear task-averaged regret can be achieved for our algorithm under mild conditions. Empirical studies on three different benchmarks clearly demonstrate the significant advantage of our algorithm over related baseline approaches. | [
"online meta-learning",
"task boundary detection",
"domain shift",
"dynamic regret",
"out of distribution detection"
] | https://openreview.net/pdf?id=LzhpSfqSXC | F9VlPVtj3G | decision | 1,700,497,716,603 | LzhpSfqSXC | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Program_Chairs"
] | decision: Accept (Oral)
title: Paper Decision |
HyOziZFh5x | NeuroMixGDP: A Neural Collapse-Inspired Random Mixup for Private Data Release | [
"Donghao Li",
"Yang Cao",
"Yuan Yao"
] | Privacy-preserving data release algorithms have gained increasing attention for their ability to protect user privacy while enabling downstream machine learning tasks. However, the utility of current popular algorithms is not always satisfactory. Mixup of raw data provides a new way of data augmentation, which can help improve utility. However, its performance drastically deteriorates when differential privacy (DP) noise is added.
To address this issue, this paper draws inspiration from the recently observed Neural Collapse (NC) phenomenon, which states that the last layer features of a neural network concentrate on the vertices of a simplex as Equiangular Tight Frame (ETF). We propose a scheme to mixup the Neural Collapse features to exploit the ETF simplex structure and release noisy mixed features to enhance the utility of the released data. By using Gaussian Differential Privacy (GDP), we obtain an asymptotic rate for the optimal mixup degree.
To further enhance the utility and address the label collapse issue when the mixup degree is large, we propose a Hierarchical sampling method to stratify the mixup samples on a small number of classes. This method remarkably improves utility when the number of classes is large. Extensive experiments demonstrate the effectiveness of our proposed method in protecting against attacks and improving utility. In particular, our approach shows significantly improved utility compared to directly training classification networks with DPSGD on CIFAR100 and MiniImagenet datasets, highlighting the benefits of using privacy-preserving data release. We release reproducible code in https://github.com/Lidonghao1996/NeuroMixGDP. | [
"Neural Collapse",
"Differential privacy",
"Private data publishing",
"Mixup"
] | https://openreview.net/pdf?id=HyOziZFh5x | pXCdbPs865 | official_review | 1,696,708,681,958 | HyOziZFh5x | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission45/Reviewer_WntU"
] | title: see below
review: Summary:
NeuroMixGDP is a novel approach to privacy-preserving data release that proposes a new mixup scheme inspired by the Neural Collapse phenomenon. The paper discusses the challenges that can arise when using the feature mixup framework, such as the sensitivity blowup of RW-Mix and label collapse. The authors examine how Avg-Mix and HS can be used to address these issues, respectively, and how these approaches informed the development of their novel designs, NeuroMixGDP(-HS). The paper provides the asymptotically optimal mixup degree rate using GDP. The proposed method is shown to significantly enhance the utility of released data while protecting user privacy. The paper also discusses the effectiveness of the proposed method in defending against attacks, such as model inversion attack and membership inference attack. Overall, the paper presents a promising approach to privacy-preserving data release that can improve the utility of released data while protecting user privacy.
Pros:
1. Enhances the utility of released data: The proposed mixup scheme can significantly improve the utility of released data, making it more useful for downstream machine learning tasks.
2. Asymptotically optimal mixup degree rate: The paper provides the asymptotically optimal mixup degree rate using basic linear model, which can help to understand the "sweet spot" choice of mixup degree.
3. Defends against attacks: The paper demonstrates that the proposed method can defend against attacks such as model inversion attack and membership inference attack.
Cons:
1. Lacks some definitions: some core definitions are lack (utility and sensitivity), which poses certain difficulties to readers who are unfamiliar with the DP area.
2. Requires further validation: while the paper presents promising results, the approach will need to be further validated and tested in larger datasets scenarios to determine its effectiveness and practicality.
3. Lacks neural collapse reference: The related neural collapse references are very minimal. It does not cite many of the theoretical works in NC literature (e.g. [1-7])
4. Some typos: the equation of y in definition 2.1 should use index j rather than i? "While the ... m samples" in line 131-133 is not a complete sentence.
Reference:
[1] Ji, Wenlong, et al. "An unconstrained layer-peeled perspective on neural collapse." arXiv preprint arXiv:2110.02796 (2021).
[2] Zhu, Zhihui, et al. "A geometric analysis of neural collapse with unconstrained features." Advances in Neural Information Processing Systems 34 (2021): 29820-29834.
[3] Han, X. Y., Vardan Papyan, and David L. Donoho. "Neural collapse under mse loss: Proximity to and dynamics on the central path." arXiv preprint arXiv:2106.02073 (2021).
[4] Zhou, Jinxin, et al. "On the optimization landscape of neural collapse under mse loss: Global optimality with unconstrained features." International Conference on Machine Learning. PMLR, 2022.
[5] Zhou, Jinxin, et al. "Are all losses created equal: A neural collapse perspective." Advances in Neural Information Processing Systems 35 (2022): 31697-31710.
[6] Mixon, Dustin G., Hans Parshall, and Jianzong Pi. "Neural collapse with unconstrained features." arXiv preprint arXiv:2011.11619 (2020).
[7] Tirer, Tom, and Joan Bruna. "Extended unconstrained features model for exploring deep neural collapse." International Conference on Machine Learning. PMLR, 2022.
rating: 7: Good paper, accept
confidence: 3: The reviewer is fairly confident that the evaluation is correct |
HyOziZFh5x | NeuroMixGDP: A Neural Collapse-Inspired Random Mixup for Private Data Release | [
"Donghao Li",
"Yang Cao",
"Yuan Yao"
] | Privacy-preserving data release algorithms have gained increasing attention for their ability to protect user privacy while enabling downstream machine learning tasks. However, the utility of current popular algorithms is not always satisfactory. Mixup of raw data provides a new way of data augmentation, which can help improve utility. However, its performance drastically deteriorates when differential privacy (DP) noise is added.
To address this issue, this paper draws inspiration from the recently observed Neural Collapse (NC) phenomenon, which states that the last layer features of a neural network concentrate on the vertices of a simplex as Equiangular Tight Frame (ETF). We propose a scheme to mixup the Neural Collapse features to exploit the ETF simplex structure and release noisy mixed features to enhance the utility of the released data. By using Gaussian Differential Privacy (GDP), we obtain an asymptotic rate for the optimal mixup degree.
To further enhance the utility and address the label collapse issue when the mixup degree is large, we propose a Hierarchical sampling method to stratify the mixup samples on a small number of classes. This method remarkably improves utility when the number of classes is large. Extensive experiments demonstrate the effectiveness of our proposed method in protecting against attacks and improving utility. In particular, our approach shows significantly improved utility compared to directly training classification networks with DPSGD on CIFAR100 and MiniImagenet datasets, highlighting the benefits of using privacy-preserving data release. We release reproducible code in https://github.com/Lidonghao1996/NeuroMixGDP. | [
"Neural Collapse",
"Differential privacy",
"Private data publishing",
"Mixup"
] | https://openreview.net/pdf?id=HyOziZFh5x | dN8goXzwEM | official_review | 1,696,643,362,589 | HyOziZFh5x | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission45/Reviewer_3QmT"
] | title: Review of NeuroMixGDP
review: Personally, I have found the exposition of this paper quite hard to follow. I do not think it is clearly outlined in the abstract and introduction how the authors leverage insights from the NC phenomenon to improve the utility of differentially private mixup. With that said, here is a summary of my reading.
Summary:
Previously, Zhang et al. proposed RW-Mix, a data augmentation technique aimed at improving the utility of (presumably) ML algorithms. However, when noise is introduced in a differentially private context, the algorithm's performance is compromised, resulting in diminished utility (whether that is for data publishing or for further use). To address this issue, this paper introduces two main algorithms: NeuroMixGDP and NeuroMixGDP-HS. The former is a differentially private version of Avg-Mix, while the latter is a differentially private adaptation of NeuroMixGDP, incorporating hierarchical sampling.
Review:
Overall, I find the idea in this paper to be novel: using insights from label collapse to propose a new algorithm in differentially private data publishing. Furthermore, the results appear to be very promising — the proposed method seems to significantly outperform existing methods, as reported in Table 2. The main issue I have is with the clarity of the paper, which I outline below:
- It is not made clear why Poisson sampling suffers from the Label Collapse issue other than being an empirical result. Is there an intuition for this, and does this apply only to Poisson sampling or more generally? If it's specific to Poisson sampling, can we employ different sampling techniques to avoid label collapse?
- Is the linear model an appropriate choice for studying and characterizing the sweet spot of the mixup degree? Were there other models that could have been considered, and if so, what are their advantages and limitations? The authors do not clearly explain why they opted for this specific model.
- I believe that Section 5, which covers the membership attack, could provide more detail and further motivation on why this is a valuable case study.
- To my knowledge, saying that an algorithm is “\mu-GDP” is not customary. In the relevant literature, a Gaussian differentially private algorithm is generally referred to as (\epsilon, \delta)-DP.
I would be more than willing to raise my score if these questions can be well addressed. As suggested by my confidence score, there may have been some parts that I may have misunderstood.
rating: 6
confidence: 3 |
HyOziZFh5x | NeuroMixGDP: A Neural Collapse-Inspired Random Mixup for Private Data Release | [
"Donghao Li",
"Yang Cao",
"Yuan Yao"
] | Privacy-preserving data release algorithms have gained increasing attention for their ability to protect user privacy while enabling downstream machine learning tasks. However, the utility of current popular algorithms is not always satisfactory. Mixup of raw data provides a new way of data augmentation, which can help improve utility. However, its performance drastically deteriorates when differential privacy (DP) noise is added.
To address this issue, this paper draws inspiration from the recently observed Neural Collapse (NC) phenomenon, which states that the last layer features of a neural network concentrate on the vertices of a simplex as Equiangular Tight Frame (ETF). We propose a scheme to mixup the Neural Collapse features to exploit the ETF simplex structure and release noisy mixed features to enhance the utility of the released data. By using Gaussian Differential Privacy (GDP), we obtain an asymptotic rate for the optimal mixup degree.
To further enhance the utility and address the label collapse issue when the mixup degree is large, we propose a Hierarchical sampling method to stratify the mixup samples on a small number of classes. This method remarkably improves utility when the number of classes is large. Extensive experiments demonstrate the effectiveness of our proposed method in protecting against attacks and improving utility. In particular, our approach shows significantly improved utility compared to directly training classification networks with DPSGD on CIFAR100 and MiniImagenet datasets, highlighting the benefits of using privacy-preserving data release. We release reproducible code in https://github.com/Lidonghao1996/NeuroMixGDP. | [
"Neural Collapse",
"Differential privacy",
"Private data publishing",
"Mixup"
] | https://openreview.net/pdf?id=HyOziZFh5x | YZpEca0XtX | meta_review | 1,699,933,212,429 | HyOziZFh5x | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission45/Area_Chair_xNCg"
] | metareview: This paper introduces a novel method for private data release, effectively balancing data utility and privacy. The approach, inspired by the neural collapse phenomenon, addresses key challenges in existing privacy-preserving algorithms. The authors have commendably responded to reviewers' feedback, enhancing the paper's clarity and empirical validation. The promising results and novel application of neural collapse insights make this submission a significant contribution to the field. Hence, I recommend its acceptance for CPAL 2024.
recommendation: Accept (Poster)
confidence: 4: The area chair is confident but not absolutely certain |
HyOziZFh5x | NeuroMixGDP: A Neural Collapse-Inspired Random Mixup for Private Data Release | [
"Donghao Li",
"Yang Cao",
"Yuan Yao"
] | Privacy-preserving data release algorithms have gained increasing attention for their ability to protect user privacy while enabling downstream machine learning tasks. However, the utility of current popular algorithms is not always satisfactory. Mixup of raw data provides a new way of data augmentation, which can help improve utility. However, its performance drastically deteriorates when differential privacy (DP) noise is added.
To address this issue, this paper draws inspiration from the recently observed Neural Collapse (NC) phenomenon, which states that the last layer features of a neural network concentrate on the vertices of a simplex as Equiangular Tight Frame (ETF). We propose a scheme to mixup the Neural Collapse features to exploit the ETF simplex structure and release noisy mixed features to enhance the utility of the released data. By using Gaussian Differential Privacy (GDP), we obtain an asymptotic rate for the optimal mixup degree.
To further enhance the utility and address the label collapse issue when the mixup degree is large, we propose a Hierarchical sampling method to stratify the mixup samples on a small number of classes. This method remarkably improves utility when the number of classes is large. Extensive experiments demonstrate the effectiveness of our proposed method in protecting against attacks and improving utility. In particular, our approach shows significantly improved utility compared to directly training classification networks with DPSGD on CIFAR100 and MiniImagenet datasets, highlighting the benefits of using privacy-preserving data release. We release reproducible code in https://github.com/Lidonghao1996/NeuroMixGDP. | [
"Neural Collapse",
"Differential privacy",
"Private data publishing",
"Mixup"
] | https://openreview.net/pdf?id=HyOziZFh5x | OyP2Y3lW83 | decision | 1,700,429,124,157 | HyOziZFh5x | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Program_Chairs"
] | decision: Accept (Oral)
comment: All reviewers and AC agreed that the paper is of high quality. This paper introduces a novel method for private data release, effectively balancing data utility and privacy. The approach, inspired by the neural collapse phenomenon, addresses key challenges in existing privacy-preserving algorithms. The authors have commendably responded to reviewers' feedback, enhancing the paper's clarity and empirical validation. The promising results and novel application of neural collapse insights make this submission a significant contribution to the field.
The action PC chair for this paper is Qing Qu, who made the decision after carefully reading the paper as well as the comments by all reviewers and AC. The decision is agreed upon by all PC chairs.
title: Paper Decision |
HyOziZFh5x | NeuroMixGDP: A Neural Collapse-Inspired Random Mixup for Private Data Release | [
"Donghao Li",
"Yang Cao",
"Yuan Yao"
] | Privacy-preserving data release algorithms have gained increasing attention for their ability to protect user privacy while enabling downstream machine learning tasks. However, the utility of current popular algorithms is not always satisfactory. Mixup of raw data provides a new way of data augmentation, which can help improve utility. However, its performance drastically deteriorates when differential privacy (DP) noise is added.
To address this issue, this paper draws inspiration from the recently observed Neural Collapse (NC) phenomenon, which states that the last layer features of a neural network concentrate on the vertices of a simplex as Equiangular Tight Frame (ETF). We propose a scheme to mixup the Neural Collapse features to exploit the ETF simplex structure and release noisy mixed features to enhance the utility of the released data. By using Gaussian Differential Privacy (GDP), we obtain an asymptotic rate for the optimal mixup degree.
To further enhance the utility and address the label collapse issue when the mixup degree is large, we propose a Hierarchical sampling method to stratify the mixup samples on a small number of classes. This method remarkably improves utility when the number of classes is large. Extensive experiments demonstrate the effectiveness of our proposed method in protecting against attacks and improving utility. In particular, our approach shows significantly improved utility compared to directly training classification networks with DPSGD on CIFAR100 and MiniImagenet datasets, highlighting the benefits of using privacy-preserving data release. We release reproducible code in https://github.com/Lidonghao1996/NeuroMixGDP. | [
"Neural Collapse",
"Differential privacy",
"Private data publishing",
"Mixup"
] | https://openreview.net/pdf?id=HyOziZFh5x | EvVfoXsh4S | official_review | 1,696,705,929,811 | HyOziZFh5x | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission45/Reviewer_T1Dw"
] | title: A new differential privacy approach inspired from Neural Collapse
review: This paper aims to mix up the Neural Collapse features to enhance the utility of the privacy-preserving data. The research topic of privacy preserving is crucial for the ML community. The main idea underlying NeuroMixGDP is instead of mixuping input features, the mixture of output features should make more sense, as the former may suffer from the broken 'Non-Approximate Collinearity (NAC)' condition. Their preliminary experiments verify NC indeed induces NAC and provide empirical evidence to support the main claim of this paper. Two technical contributions, Averaging Mixup and Hierarchical Sampling, were proposed to significantly boost the performance of NeuroMixGDP. Experimental results also verify the effectiveness of NeuroMixGDP.
Strengthness:
(1) The paper is well-motivated and supported by results.
(2) I like the preliminary results provided in Figure 1 and Figure 2, which provides clear evidence and important insights to readers.
(3) The authors also provide results of Model Inversion Attack, making the paper more solid.
Weakness:
(1) Can the authors elaborate more the experimental settings of Figure 1 and Figure 2. I would like to see the similar trend on various settings and therefore, the finding in this paper is a general one that can be observed across various settings.
(2) Are we solely extracting the output features of the last layer or all layers? I expect only the features from last layer are helpful here.
(3) What does epsilon in Figure 2 stand for?
(4) I encourage the authors to explain the Poisson Sampling in the early part of the paper.
(5) what is the formulation or definition of sensitivity mean here? what its relationship to DP noise?
rating: 7: Good paper, accept
confidence: 3: The reviewer is fairly confident that the evaluation is correct |
H2rCZCfXkS | Jaxpruner: A Concise Library for Sparsity Research | [
"Joo Hyung Lee",
"Wonpyo Park",
"Nicole Elyse Mitchell",
"Jonathan Pilault",
"Johan Samir Obando Ceron",
"Han-Byul Kim",
"Namhoon Lee",
"Elias Frantar",
"Yun Long",
"Amir Yazdanbakhsh",
"Woohyun Han",
"Shivani Agrawal",
"Suvinay Subramanian",
"Xin Wang",
"Sheng-Chun Kao",
"Xingyao Zhang",
"Trevor Gale",
"Aart J.C. Bik",
"Milen Ferev",
"Zhonglin Han",
"Hong-Seok Kim",
"Yann Dauphin",
"Gintare Karolina Dziugaite",
"Pablo Samuel Castro",
"Utku Evci"
] | This paper introduces JaxPruner, an open-source JAX-based pruning and sparse training library for machine learning research. JaxPruner aims to accelerate research on sparse neural networks by providing concise implementations of popular pruning and sparse training algorithms with minimal memory and latency overhead. Algorithms implemented in JaxPruner use a common API and work seamlessly with the popular optimization library Optax, which, in turn, enables easy integration with existing JAX based libraries. We demonstrate this ease of integration by providing examples in four different codebases: Scenic, t5x, Dopamine and FedJAX and provide baseline experiments on popular benchmarks. Jaxpruner is hosted at github.com/google-research/jaxpruner | [
"jax",
"sparsity",
"pruning",
"quantization",
"sparse training",
"efficiency",
"library",
"software"
] | https://openreview.net/pdf?id=H2rCZCfXkS | T3D1grJbpP | official_review | 1,696,692,218,248 | H2rCZCfXkS | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission14/Reviewer_Kqti"
] | title: Official Review of Submission 14
review: ## Overview
The authors introduce a new sparsity-focused library tailored for machine learning research. Built on the JAX platform, the library facilitates the implementation of an array of sparse algorithms, encompassing various sparsity patterns, both static and dynamic training sparsity, and multiple pruning metrics. A comprehensive set of experiments spanning diverse domains such as image recognition, natural language processing, federated learning, and deep reinforcement learning, attests to the library's adaptability.
## Comments
(+) The library has been effectively tested on an array of tasks, from image recognition to deep reinforcement learning, and incorporates numerous sparsity methodologies, from unstructured to dynamic sparsity. This demonstrates the library's comprehensive adaptability.
(+) Some features, like using int8 type for storing masks, are commendable, particularly given the growing prominence of expansive foundation models that demand extensive mask memory.
(+) The paper is well organized, delving initially into the reasons for creating a new sparsity-centered library, transitioning into its primary features, and culminating with in-depth benchmarking results.
(+) This library makes a notable contribution to the sparsity community, simplifying the process of implementing new ideas.
(-) An elaboration on its compatibility with other sparsity training models, like Mixture-of-Experts, would be beneficial.
(-) The paper lacks a comparative analysis with prevalent libraries (w. Pytorch or Tensorflow), particularly concerning training/inference velocity and memory consumption.
rating: 8: Top 50% of accepted papers, clear accept
confidence: 4: The reviewer is confident but not absolutely certain that the evaluation is correct |
H2rCZCfXkS | Jaxpruner: A Concise Library for Sparsity Research | [
"Joo Hyung Lee",
"Wonpyo Park",
"Nicole Elyse Mitchell",
"Jonathan Pilault",
"Johan Samir Obando Ceron",
"Han-Byul Kim",
"Namhoon Lee",
"Elias Frantar",
"Yun Long",
"Amir Yazdanbakhsh",
"Woohyun Han",
"Shivani Agrawal",
"Suvinay Subramanian",
"Xin Wang",
"Sheng-Chun Kao",
"Xingyao Zhang",
"Trevor Gale",
"Aart J.C. Bik",
"Milen Ferev",
"Zhonglin Han",
"Hong-Seok Kim",
"Yann Dauphin",
"Gintare Karolina Dziugaite",
"Pablo Samuel Castro",
"Utku Evci"
] | This paper introduces JaxPruner, an open-source JAX-based pruning and sparse training library for machine learning research. JaxPruner aims to accelerate research on sparse neural networks by providing concise implementations of popular pruning and sparse training algorithms with minimal memory and latency overhead. Algorithms implemented in JaxPruner use a common API and work seamlessly with the popular optimization library Optax, which, in turn, enables easy integration with existing JAX based libraries. We demonstrate this ease of integration by providing examples in four different codebases: Scenic, t5x, Dopamine and FedJAX and provide baseline experiments on popular benchmarks. Jaxpruner is hosted at github.com/google-research/jaxpruner | [
"jax",
"sparsity",
"pruning",
"quantization",
"sparse training",
"efficiency",
"library",
"software"
] | https://openreview.net/pdf?id=H2rCZCfXkS | 8dHipclHly | official_review | 1,696,649,440,949 | H2rCZCfXkS | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission14/Reviewer_m57v"
] | title: Good Library of Jax for Network Pruning
review: This paper presents a new neural network pruning library implemented with Jax, called JaxPruner, with 3 tenets: (a) fast integration (b) research first (c) minimal overhead. The library has implemented the major components of network pruning, supporting different pruning criteria, schedules, etc. They also establish the results of many baseline pruning schemes. This work does not introduce new pruning algorithms. Instead, they offer a new pruning library with Jax, which could benefit the Jax & pruning community.
Pros:
1. First and foremost, there is no prevailing Jax library for neural network pruning. This paper bridges this gap, which could benefit many researchers, esp. those using Jax as their deep learning framework and are interested in pruning.
2. JaxPruner is featured by fast integration, research first, and minimal overhead, which could make the library easy to use.
3. JaxPruner also provides strong baselines in either traditional pruning (pruning a pretrained model) or sparse training (pruning at initialization).
Cons:
1. Some of the results look interesting but lack enough discussion. E.g., in Tab. 1, for the row ViT-B16+, SET and RigL beat Dense with the prolonged training, while underperforming Dense in the row ViT-B16. This is quite unusual. Did the authors double-check and confirm the results are correct? If so, why? More explanations or discussions are highly suggested. Unreliable baseline results could mislead the community.
2. The library implemented many baseline pruning schemes like RAND, SAL, and MAG -- these are different pruning criteria. There is another major group of pruning methods that use sparsity-inducing penalty terms for sparsity, such as [*1-*4]. I was wondering whether this group of methods can easily fit into JaxPruner?
3. Typos: Line 258. with block sparsit; -> sparsity.
- [*1] 2016-NeurIPS-Learning Structured Sparsity in Deep Neural Networks
- [*2] 2018-ICLR-Learning Sparse Neural Networks through L0 Regularization
- [*3] 2021-ICLR-Neural Pruning via Growing Regularization
- [*4] 2022-ICLR-Dual Lottery Ticket Hypothesis
rating: 6: Marginally above acceptance threshold
confidence: 5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature |
H2rCZCfXkS | Jaxpruner: A Concise Library for Sparsity Research | [
"Joo Hyung Lee",
"Wonpyo Park",
"Nicole Elyse Mitchell",
"Jonathan Pilault",
"Johan Samir Obando Ceron",
"Han-Byul Kim",
"Namhoon Lee",
"Elias Frantar",
"Yun Long",
"Amir Yazdanbakhsh",
"Woohyun Han",
"Shivani Agrawal",
"Suvinay Subramanian",
"Xin Wang",
"Sheng-Chun Kao",
"Xingyao Zhang",
"Trevor Gale",
"Aart J.C. Bik",
"Milen Ferev",
"Zhonglin Han",
"Hong-Seok Kim",
"Yann Dauphin",
"Gintare Karolina Dziugaite",
"Pablo Samuel Castro",
"Utku Evci"
] | This paper introduces JaxPruner, an open-source JAX-based pruning and sparse training library for machine learning research. JaxPruner aims to accelerate research on sparse neural networks by providing concise implementations of popular pruning and sparse training algorithms with minimal memory and latency overhead. Algorithms implemented in JaxPruner use a common API and work seamlessly with the popular optimization library Optax, which, in turn, enables easy integration with existing JAX based libraries. We demonstrate this ease of integration by providing examples in four different codebases: Scenic, t5x, Dopamine and FedJAX and provide baseline experiments on popular benchmarks. Jaxpruner is hosted at github.com/google-research/jaxpruner | [
"jax",
"sparsity",
"pruning",
"quantization",
"sparse training",
"efficiency",
"library",
"software"
] | https://openreview.net/pdf?id=H2rCZCfXkS | 5SoXi1Psxm | official_review | 1,696,640,176,260 | H2rCZCfXkS | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission14/Reviewer_xdL7"
] | title: Official Review for JAXpruner: An open-source JAX-based pruning and sparse training library
review: The paper presents a JAX-based sparsity library for machine learning research. JaxPruner provide concise implementations of popular pruning and sparse training algorithms with minimal memory and latency overhead. The Optax gradient transformations for implementing algorithms is interesting. The paper is well written with appropriate code snippets to elucidate usage. Minimization of memory and run-time overhead by compressing binary masks for representing sparsity and incorporation of N:M sparsity are important features to have. However, I still feel the coverage of various popular pruning baseline algorithms are still missing (eg. SNIP, GrasP, SynFlow etc) if the focus is research. I find the further experiments section very interesting and informative. Some typos need to be fixed immediately (eg. Section 6- Concusion -> Conclusion). Overall, it is an important tool for sparse community and I recommend acceptance.
rating: 7: Good paper, accept
confidence: 5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature |
H2rCZCfXkS | Jaxpruner: A Concise Library for Sparsity Research | [
"Joo Hyung Lee",
"Wonpyo Park",
"Nicole Elyse Mitchell",
"Jonathan Pilault",
"Johan Samir Obando Ceron",
"Han-Byul Kim",
"Namhoon Lee",
"Elias Frantar",
"Yun Long",
"Amir Yazdanbakhsh",
"Woohyun Han",
"Shivani Agrawal",
"Suvinay Subramanian",
"Xin Wang",
"Sheng-Chun Kao",
"Xingyao Zhang",
"Trevor Gale",
"Aart J.C. Bik",
"Milen Ferev",
"Zhonglin Han",
"Hong-Seok Kim",
"Yann Dauphin",
"Gintare Karolina Dziugaite",
"Pablo Samuel Castro",
"Utku Evci"
] | This paper introduces JaxPruner, an open-source JAX-based pruning and sparse training library for machine learning research. JaxPruner aims to accelerate research on sparse neural networks by providing concise implementations of popular pruning and sparse training algorithms with minimal memory and latency overhead. Algorithms implemented in JaxPruner use a common API and work seamlessly with the popular optimization library Optax, which, in turn, enables easy integration with existing JAX based libraries. We demonstrate this ease of integration by providing examples in four different codebases: Scenic, t5x, Dopamine and FedJAX and provide baseline experiments on popular benchmarks. Jaxpruner is hosted at github.com/google-research/jaxpruner | [
"jax",
"sparsity",
"pruning",
"quantization",
"sparse training",
"efficiency",
"library",
"software"
] | https://openreview.net/pdf?id=H2rCZCfXkS | 5Hu7ksbjKI | meta_review | 1,699,829,727,880 | H2rCZCfXkS | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission14/Area_Chair_EENM"
] | metareview: This submission presents a sparsity-centric JAX library designed for machine learning research. Leveraging the JAX platform, this library streamlines the development of a wide range of sparse algorithms, accommodating diverse sparsity patterns, including both static and dynamically evolving training sparsity, as well as incorporating multiple pruning metrics. All authors voted for the acceptance. I also see it will benefit the sparsity community a lot.
recommendation: Accept (Oral)
confidence: 4: The area chair is confident but not absolutely certain |
H2rCZCfXkS | Jaxpruner: A Concise Library for Sparsity Research | [
"Joo Hyung Lee",
"Wonpyo Park",
"Nicole Elyse Mitchell",
"Jonathan Pilault",
"Johan Samir Obando Ceron",
"Han-Byul Kim",
"Namhoon Lee",
"Elias Frantar",
"Yun Long",
"Amir Yazdanbakhsh",
"Woohyun Han",
"Shivani Agrawal",
"Suvinay Subramanian",
"Xin Wang",
"Sheng-Chun Kao",
"Xingyao Zhang",
"Trevor Gale",
"Aart J.C. Bik",
"Milen Ferev",
"Zhonglin Han",
"Hong-Seok Kim",
"Yann Dauphin",
"Gintare Karolina Dziugaite",
"Pablo Samuel Castro",
"Utku Evci"
] | This paper introduces JaxPruner, an open-source JAX-based pruning and sparse training library for machine learning research. JaxPruner aims to accelerate research on sparse neural networks by providing concise implementations of popular pruning and sparse training algorithms with minimal memory and latency overhead. Algorithms implemented in JaxPruner use a common API and work seamlessly with the popular optimization library Optax, which, in turn, enables easy integration with existing JAX based libraries. We demonstrate this ease of integration by providing examples in four different codebases: Scenic, t5x, Dopamine and FedJAX and provide baseline experiments on popular benchmarks. Jaxpruner is hosted at github.com/google-research/jaxpruner | [
"jax",
"sparsity",
"pruning",
"quantization",
"sparse training",
"efficiency",
"library",
"software"
] | https://openreview.net/pdf?id=H2rCZCfXkS | 55gyFDrnvc | decision | 1,700,395,709,784 | H2rCZCfXkS | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Program_Chairs"
] | decision: Accept (Oral)
comment: The paper introduces JaxPruner, a JAX-based sparsity library for machine learning research, focusing on neural network pruning. It provides concise implementations of various pruning and sparse training algorithms, with minimal memory and latency overhead. Reviewers commend the library's comprehensive adaptability and features, such as using int8 type for storing masks, which is particularly useful for large foundation models. However, they suggest providing more information on compatibility with other sparsity training models and conducting a comparative analysis with prevalent libraries like PyTorch or TensorFlow, especially concerning training/inference velocity and memory consumption. Additionally, one reviewer raises questions about the reliability of certain baseline results and suggests more detailed explanations or discussions for unusual findings.
The action PC chair for this paper is Atlas Wang, who made the decision after carefully reading the paper as well as the comments by all reviewers and AC. The decision is agreed by all PC chairs.
title: Paper Decision |
GLQJH92m1m | Image Quality Assessment: Integrating Model-centric and Data-centric Approaches | [
"Peibei Cao",
"Dingquan Li",
"Kede Ma"
] | Learning-based image quality assessment (IQA) has made remarkable progress in the past decade, but nearly all consider the two key components---model and data---in relative isolation. Specifically, model-centric IQA focuses on developing "better" objective quality methods on fixed and extensively reused datasets, with a great danger of overfitting. Data-centric IQA involves conducting psychophysical experiments to construct "better" human-annotated datasets, which unfortunately ignores current IQA models during dataset creation. In this paper, we first design a series of experiments to probe computationally that such isolation of model and data impedes further progress of IQA. We then describe a computational framework that integrates model-centric and data-centric IQA. As a specific example, we design computational modules to quantify the sampling-worthiness of candidate images based on blind IQA (BIQA) model predictions and deep content-aware features. Experimental results show that the proposed sampling-worthiness module successfully spots diverse failures of the examined BIQA models, which are indeed worthy samples to be included in next-generation datasets. | [
"Learning-based IQA",
"model-centric IQA",
"data-centric IQA",
"sampling-worthiness."
] | https://openreview.net/pdf?id=GLQJH92m1m | r6nCiwYCwO | official_review | 1,694,572,372,736 | GLQJH92m1m | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission3/Reviewer_5DQU"
] | title: Simple but effective idea
review: This paper presents a new IQA method that combines the data-centric IQA and model-centric IQA. Basically, this paper is more likely an active learning method in my opinion. Using existing IQA models and VGGNet to select the most valuable data.
At first, the author demonstrates the drawbacks of the previous data-centric IQA and model-centric IQA. Data-centric suffers from the easy dataset problem and model-centric suffers from the overfitting problem. Then introduce a method that uses IQA models and VGGNet to find the difficult and diverse samples.
The reason for the improvement of this method also seems to be intuitive. There are no obvious weaknesses in this paper to me.
I am not familiar with this field but I do buy the main idea of this paper.
rating: 6: Marginally above acceptance threshold
confidence: 1: The reviewer's evaluation is an educated guess |
GLQJH92m1m | Image Quality Assessment: Integrating Model-centric and Data-centric Approaches | [
"Peibei Cao",
"Dingquan Li",
"Kede Ma"
] | Learning-based image quality assessment (IQA) has made remarkable progress in the past decade, but nearly all consider the two key components---model and data---in relative isolation. Specifically, model-centric IQA focuses on developing "better" objective quality methods on fixed and extensively reused datasets, with a great danger of overfitting. Data-centric IQA involves conducting psychophysical experiments to construct "better" human-annotated datasets, which unfortunately ignores current IQA models during dataset creation. In this paper, we first design a series of experiments to probe computationally that such isolation of model and data impedes further progress of IQA. We then describe a computational framework that integrates model-centric and data-centric IQA. As a specific example, we design computational modules to quantify the sampling-worthiness of candidate images based on blind IQA (BIQA) model predictions and deep content-aware features. Experimental results show that the proposed sampling-worthiness module successfully spots diverse failures of the examined BIQA models, which are indeed worthy samples to be included in next-generation datasets. | [
"Learning-based IQA",
"model-centric IQA",
"data-centric IQA",
"sampling-worthiness."
] | https://openreview.net/pdf?id=GLQJH92m1m | nGflXoudqw | decision | 1,700,363,432,816 | GLQJH92m1m | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Program_Chairs"
] | decision: Accept (Oral)
comment: The paper addresses the problem of image quality assessment (IQA) and introduces a computational framework that integrates model-centric and data-centric IQA, to quantify the sampling-worthiness of candidate images based on blind IQA (BIQA) - hence clicking the data efficiency theme of CPAL. Reviewers appreciate the paper's contributions in recognizing the overfitting issue and the problem of simple datasets in blind IQA. They find the idea of integrating both approaches promising and interesting for future advancements in the field. However, concerns are raised about the writing style and clarity of the paper, suggesting that it should provide more background information and rationale for certain choices. Additionally, one reviewer suggests considering "conditioned" evaluations in addition to "unconditioned" ones. Overall, the paper's main idea and contributions are well-received by the reviewers.
The action PC chair for this paper is Atlas Wang, who made the decision after carefully reading the paper as well as the comments by all reviewers and AC. The decision is agreed by all PC chairs.
title: Paper Decision |
GLQJH92m1m | Image Quality Assessment: Integrating Model-centric and Data-centric Approaches | [
"Peibei Cao",
"Dingquan Li",
"Kede Ma"
] | Learning-based image quality assessment (IQA) has made remarkable progress in the past decade, but nearly all consider the two key components---model and data---in relative isolation. Specifically, model-centric IQA focuses on developing "better" objective quality methods on fixed and extensively reused datasets, with a great danger of overfitting. Data-centric IQA involves conducting psychophysical experiments to construct "better" human-annotated datasets, which unfortunately ignores current IQA models during dataset creation. In this paper, we first design a series of experiments to probe computationally that such isolation of model and data impedes further progress of IQA. We then describe a computational framework that integrates model-centric and data-centric IQA. As a specific example, we design computational modules to quantify the sampling-worthiness of candidate images based on blind IQA (BIQA) model predictions and deep content-aware features. Experimental results show that the proposed sampling-worthiness module successfully spots diverse failures of the examined BIQA models, which are indeed worthy samples to be included in next-generation datasets. | [
"Learning-based IQA",
"model-centric IQA",
"data-centric IQA",
"sampling-worthiness."
] | https://openreview.net/pdf?id=GLQJH92m1m | fogBvT6EBp | official_review | 1,696,886,722,368 | GLQJH92m1m | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission3/Reviewer_fMGt"
] | title: Review for Submission #3
review: **Paper Summary:**
This paper identifies the overfitting issue and the problem of simple datasets in blind IQA, possibly resulting from the isolation of the IQA model and data. In response, the paper presents a computational framework that integrates model-centric and data-centric IQA. This is achieved by enhancing the quality predictor with an auxiliary module to guide the sampling process. Its effectiveness is substantiated through a specific demonstration.
**Pros:**
1. Integrating model-centric and data-centric IQA is a promising approach and may shed light on future advancements in this field.
2. Recognizing the overfitting issue and the easy dataset problem can serve as valuable references for the community.
**Cons:**
1. A primary drawback of this work is its writing style, which may prove challenging for general readers unfamiliar with IQA. Ideally, the authors should initiate with the underlying rationales before delving into experimental details. The current presentation makes it challenging to grasp the nature of the easy dataset problem, and statements like "the newly created ones are more difficult to challenge the most recent UNIQUE" remain vague. Furthermore, certain background information and details, such as the outputs of an IQA model and the rationale behind the metrics chosen in Section 2, are essential for comprehending the paper's core content. The authors are strongly encouraged to refine their writing for broader accessibility.
2. Given that this paper's main technical contribution focuses on utilizing existing IQA models to inform the development of new IQA datasets, the phrase "integrating model-centric and data-centric IQA" might not accurately reflect the technical contribution of this work.
rating: 6: Marginally above acceptance threshold
confidence: 2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper |
GLQJH92m1m | Image Quality Assessment: Integrating Model-centric and Data-centric Approaches | [
"Peibei Cao",
"Dingquan Li",
"Kede Ma"
] | Learning-based image quality assessment (IQA) has made remarkable progress in the past decade, but nearly all consider the two key components---model and data---in relative isolation. Specifically, model-centric IQA focuses on developing "better" objective quality methods on fixed and extensively reused datasets, with a great danger of overfitting. Data-centric IQA involves conducting psychophysical experiments to construct "better" human-annotated datasets, which unfortunately ignores current IQA models during dataset creation. In this paper, we first design a series of experiments to probe computationally that such isolation of model and data impedes further progress of IQA. We then describe a computational framework that integrates model-centric and data-centric IQA. As a specific example, we design computational modules to quantify the sampling-worthiness of candidate images based on blind IQA (BIQA) model predictions and deep content-aware features. Experimental results show that the proposed sampling-worthiness module successfully spots diverse failures of the examined BIQA models, which are indeed worthy samples to be included in next-generation datasets. | [
"Learning-based IQA",
"model-centric IQA",
"data-centric IQA",
"sampling-worthiness."
] | https://openreview.net/pdf?id=GLQJH92m1m | OLsjDqxPle | official_review | 1,696,851,410,643 | GLQJH92m1m | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission3/Reviewer_1cbT"
] | title: Interesting study and discussions
review: This paper discusses the important problem of image quality assessment (IQA), which is essential to the evaluation of many computer vision applications. It brings forward that the isolation of model-centric and data-centric approaches impedes further progress in IQA, which is not well-addressed in previous works. The argument is supported by studies of existing IQA methods. It also proposes a novel IQA framework integrating model-centric and data-centric methods, which shows superiority compared to past works. The intuitions and discussions in this paper are interesting and motivating, and I believe they can call for more attention to this important problem as well as inspire more future research.
This paper mainly focuses on "unconditioned" evaluation of image qualities, but it'd be more interesting to also discuss "conditioned" evaluations, as they are widely used in a lot of tasks like image generation, novel view synthesis (e.g. NeRF), etc. For example, data-based metrics such as FID, KID, and model-based metrics such as PSNR, SSIM, LPIPS.
rating: 7: Good paper, accept
confidence: 3: The reviewer is fairly confident that the evaluation is correct |
GLQJH92m1m | Image Quality Assessment: Integrating Model-centric and Data-centric Approaches | [
"Peibei Cao",
"Dingquan Li",
"Kede Ma"
] | Learning-based image quality assessment (IQA) has made remarkable progress in the past decade, but nearly all consider the two key components---model and data---in relative isolation. Specifically, model-centric IQA focuses on developing "better" objective quality methods on fixed and extensively reused datasets, with a great danger of overfitting. Data-centric IQA involves conducting psychophysical experiments to construct "better" human-annotated datasets, which unfortunately ignores current IQA models during dataset creation. In this paper, we first design a series of experiments to probe computationally that such isolation of model and data impedes further progress of IQA. We then describe a computational framework that integrates model-centric and data-centric IQA. As a specific example, we design computational modules to quantify the sampling-worthiness of candidate images based on blind IQA (BIQA) model predictions and deep content-aware features. Experimental results show that the proposed sampling-worthiness module successfully spots diverse failures of the examined BIQA models, which are indeed worthy samples to be included in next-generation datasets. | [
"Learning-based IQA",
"model-centric IQA",
"data-centric IQA",
"sampling-worthiness."
] | https://openreview.net/pdf?id=GLQJH92m1m | 9auffFGtiw | meta_review | 1,699,671,283,518 | GLQJH92m1m | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission3/Area_Chair_LPnk"
] | metareview: The paper proposes a unifying approach that combines data-centric and model-centric methods for image quality assessment (IQA) problems. The authors should revise the paper to improve the readability. All reviewers agree to accept the paper.
recommendation: Accept (Poster)
confidence: 4: The area chair is confident but not absolutely certain |
FG8b2I2AkF | How to Prune Your Language Model: Recovering Accuracy on the ``Sparsity May Cry'' Benchmark | [
"Eldar Kurtic",
"Torsten Hoefler",
"Dan Alistarh"
] | Pruning large language models (LLMs) from the BERT family has emerged as a standard compression benchmark, and several pruning methods have been proposed for this task.
The recent ``Sparsity May Cry'' (SMC) benchmark put into question the validity of all existing methods, exhibiting a more complex setup where many known pruning methods appear to fail.
We revisit the question of accurate BERT-pruning during fine-tuning on downstream datasets, and propose a set of general guidelines for successful pruning, even on the challenging SMC benchmark.
First, we perform a cost-vs-benefits analysis of pruning model components, such as the embeddings and the classification head; second, we provide a simple-yet-general way of scaling training, sparsification and learning rate schedules relative to the desired target sparsity; finally, we investigate the importance of proper parametrization for Knowledge Distillation in the context of LLMs. Our simple insights lead to state-of-the-art results, both on classic BERT-pruning benchmarks, as well as on the SMC benchmark, showing that even classic gradual magnitude pruning (GMP) can yield competitive results, with the right approach. | [
"pruning",
"deep learning",
"benchmarking"
] | https://openreview.net/pdf?id=FG8b2I2AkF | go9dOK4148 | meta_review | 1,699,636,005,965 | FG8b2I2AkF | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission52/Area_Chair_JstC"
] | metareview: (this is an invited paper, so only one meta-review is provided)
The paper addresses the crucial topic of unstructured pruning in language models (LMs) with a particular focus on the recently proposed "Sparsity May Cry (SMC-Bench)" benchmark. The authors respond to SMC-Bench's surprising conclusion that state-of-the-art (SOTA) sparse algorithms fail consistently, even at low sparsity levels. SMC poses a new challenging benchmark for the community, especially for large-scale settings. This paper comprehensively studies the BERT pruning and provides a set of general guidelines for successful pruning on SMC-Bench. The main message delivered by this paper is that following a set of three guidelines, simple gradual magnitude pruning (GMP) can yield competitive results. Overall, the authors have done an excellent job of summarizing the principles for BERT-pruning. The contribution of this paper is timely and significant for the community.
Contributions and Strengths:
1. Pruning Best Practices: The authors make a significant contribution by proposing a set of pruning best practices for BERT-pruning. These guidelines, presented for the first time or refined from existing literature, include considerations for the post-pruning training period, sparsification components, learning rate schedules, and knowledge distillation.
2. Deconstruction of SMC-Bench: The paper meticulously analyzes the relationship between the proposed best practices and the SMC-Bench benchmark. By applying their guidelines to the benchmark's setup, the authors challenge the negative claims made by SMC-Bench, even for basic gradual magnitude pruning (GMP). The authors support their claims with concrete and convincing results.
3. In-depth Analysis: I like the analysis of the efficiency analysis of different components in RoBERTa, which provides convincing and clear guidance for practitioners and researchers. The extensive ablation study in Table 2-5 is also rich and useful.
4. Performance on Hardest Tasks: The authors demonstrate the effectiveness of their proposed guidelines on the tasks identified by SMC-Bench as the "hardest" – CommonsenseQA and WinoGrande. Across both settings, their approach outperforms existing results on the benchmark, achieving high sparsities of 80-90% accurately.
5. State-of-the-Art Results: The paper claims new state-of-the-art sparsity-vs-accuracy results for the BERT-base model on the SQuADv1.1 task. Moreover, their approach, coupled with the oBERT pruner, achieves stable results when pruning RoBERTa-large up to 90% sparsity, challenging the notion that unstructured sparsity is not viable in challenging settings.
Overall, the paper makes a valuable contribution to the field of BERT-pruning by proposing effective best practices and challenging the findings of the SMC-Bench benchmark. This paper, together with SMC-Bench, provides a clearer picture of whether pruning a larger language model is easier or harder, summarizing the overall progress of pruning in the context of BERT. It would be a valuable and welcome contribution to CPAL.
recommendation: Accept (Oral)
confidence: 5: The area chair is absolutely certain |
FG8b2I2AkF | How to Prune Your Language Model: Recovering Accuracy on the ``Sparsity May Cry'' Benchmark | [
"Eldar Kurtic",
"Torsten Hoefler",
"Dan Alistarh"
] | Pruning large language models (LLMs) from the BERT family has emerged as a standard compression benchmark, and several pruning methods have been proposed for this task.
The recent ``Sparsity May Cry'' (SMC) benchmark put into question the validity of all existing methods, exhibiting a more complex setup where many known pruning methods appear to fail.
We revisit the question of accurate BERT-pruning during fine-tuning on downstream datasets, and propose a set of general guidelines for successful pruning, even on the challenging SMC benchmark.
First, we perform a cost-vs-benefits analysis of pruning model components, such as the embeddings and the classification head; second, we provide a simple-yet-general way of scaling training, sparsification and learning rate schedules relative to the desired target sparsity; finally, we investigate the importance of proper parametrization for Knowledge Distillation in the context of LLMs. Our simple insights lead to state-of-the-art results, both on classic BERT-pruning benchmarks, as well as on the SMC benchmark, showing that even classic gradual magnitude pruning (GMP) can yield competitive results, with the right approach. | [
"pruning",
"deep learning",
"benchmarking"
] | https://openreview.net/pdf?id=FG8b2I2AkF | 8hlR2u6hzh | decision | 1,700,497,724,762 | FG8b2I2AkF | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Program_Chairs"
] | decision: Accept (Oral)
title: Paper Decision |
D9ggc3l0wi | Leveraging Sparse Input and Sparse Models: Efficient Distributed Learning in Resource-Constrained Environments | [
"Emmanouil Kariotakis",
"Grigorios Tsagkatakis",
"Panagiotis Tsakalides",
"Anastasios Kyrillidis"
] | Optimizing for reduced computational and bandwidth resources enables model training in less-than-ideal environments and paves the way for practical and accessible AI solutions. This work is about the study and design of a system that exploits sparsity in the input layer and intermediate layers of a neural network. Further, the system gets trained and operates in a distributed manner. Focusing on image classification tasks, our system efficiently utilizes reduced portions of the input image data. By exploiting transfer learning techniques, it employs a pre-trained feature extractor, with the encoded representations being subsequently introduced into selected subnets of the system's final classification module, adopting the Independent Subnetwork Training (IST) algorithm. This way, the input and subsequent feedforward layers are trained via sparse ``actions'', where input and intermediate features are subsampled and propagated in the forward layers.
We conduct experiments on several benchmark datasets, including CIFAR-$10$, NWPU-RESISC$45$, and the Aerial Image dataset. The results consistently showcase appealing accuracy despite sparsity: it is surprising that, empirically, there are cases where fixed masks could potentially outperform random masks and that the model achieves comparable or even superior accuracy with only a fraction ($50\%$ or less) of the original image, making it particularly relevant in bandwidth-constrained scenarios. This further highlights the robustness of learned features extracted by ViT, offering the potential for parsimonious image data representation with sparse models in distributed learning. | [
"sparse neural network training",
"efficient training"
] | https://openreview.net/pdf?id=D9ggc3l0wi | vEtDb37Auu | official_review | 1,696,924,047,708 | D9ggc3l0wi | [
"everyone"
] | [
"CPAL.cc/2024/Conference/Submission46/Reviewer_Dpaw"
] | title: Review
review: Pros:
---
- The writing is clear and the paper is well-structured.
- The idea presented is novel, leveraging both sparse input and sparse models in the distributed training setting to further enhance efficiency.
- The experiments are sufficient.
Cons/Other Comments:
---
- I'm curious to know: if we sparsify and train the entire network instead of using a pre-trained ViT as feature extractors, will the proposed method still be effective?
- Perhaps I overlooked something, but I'm unclear on how to choose the fixed mask.
- With only 25% of parameters in a 4-worker setting, there appears to be a significant performance drop.
- It would strengthen the argument if the authors could provide actual training times in the distributed setting with varying worker counts.
rating: 6: Marginally above acceptance threshold
confidence: 4: The reviewer is confident but not absolutely certain that the evaluation is correct |
Subsets and Splits