image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
Fig. 1:Shape denoising example. The noisy shape (b) has been obtained from the original shape (a) by a noise inducing process such as those described in Section2.2. A shape denoising method is used to obtain the denoised shape (c). | 2402.15587v1 | 1 |
|
Fig. 2:Illustration of shape alignment and the six types of shape noise introduced in this work. (a) Shape before alignment, (b) Shape after alignment, (c) Salt and pepper noise, (d) Circle noise, (e) Real image noise, (f) Occlusion noise, (g) Thresholded probability noise, (h) Detection image noise. | 2402.15587v1 | 2 |
|
Figure 1:Distilling adversarial robustness using heterogeneous teachers (DARHT).Adversarial examples are generated by applying an adversarial training algorithm to the clean examples using the student model.
Subsequently, DARHT selects adversarial or clean examples as input for both the student model and heterogeneous teachers.
The knowledge distillation loss (ℒ𝒦subscriptℒ𝒦\mathcal{L}_{\mathcal{K}}caligraphic_L start_POSTSUBSCRIPT caligraphic_K end_POSTSUBSCRIPT) is calculated using the teacher logits and the student-teacher feature map.
The classification loss (ℒ𝒞subscriptℒ𝒞\mathcal{L}_{\mathcal{C}}caligraphic_L start_POSTSUBSCRIPT caligraphic_C end_POSTSUBSCRIPT) is computed as the cross-entropy between the student logits and the labels. | 2402.15586v1 | 1 |
|
Figure 2:Distilling knowledge from teacher logits to student-teacher feature map.An example (x𝑥xitalic_x) passed to the student (Sθsubscript𝑆𝜃S_{\theta}italic_S start_POSTSUBSCRIPT italic_θ end_POSTSUBSCRIPT) and three teachers (T1subscript𝑇1T_{1}italic_T start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT,T2subscript𝑇2T_{2}italic_T start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT, andT3subscript𝑇3T_{3}italic_T start_POSTSUBSCRIPT 3 end_POSTSUBSCRIPT) models. | 2402.15586v1 | 2 |
|
Figure 3:The performance of the student (ResNet-18) on training accuracy and W-Robust metrics under FGSM, PGD, and CW∞{}_{\infty}start_FLOATSUBSCRIPT ∞ end_FLOATSUBSCRIPTattacks is evaluated with three distinct groups of teachers: Homo-ResNet, Hetero-Mix, and Homo-ViT on the CIFAR-100 dataset. | 2402.15586v1 | 3 |
|
Figure 4:APGD Robustness and clean accuracies for the DARHT student (ResNet-18) and different teacher compositions across epochs on CIFAR-100. | 2402.15586v1 | 4 |
|
Figure 5:W-Robust distributions with respect to the APGD and Square attacks for MTARD, RSLAD and DARHT. | 2402.15586v1 | 5 |
|
Figure 6:DARHT performance across epochs for heterogeneous adversarial training methods.Average robustness (left) and clean accuracy (right) were tracked across epochs 60 to 70 for DARHT configurations with teachers trained using the same (FAT, red) or different (TRADES and FAT, blue) adversarial training methods. | 2402.15586v1 | 6 |
|
(a)interlaken_00_b𝑖𝑛𝑡𝑒𝑟𝑙𝑎𝑘𝑒𝑛_00_𝑏interlaken\_00\_bitalic_i italic_n italic_t italic_e italic_r italic_l italic_a italic_k italic_e italic_n _ 00 _ italic_b | 2402.15584v3 | 5 |
|
(a)interlaken_00_b𝑖𝑛𝑡𝑒𝑟𝑙𝑎𝑘𝑒𝑛_00_𝑏interlaken\_00\_bitalic_i italic_n italic_t italic_e italic_r italic_l italic_a italic_k italic_e italic_n _ 00 _ italic_b | 2402.15584v3 | 6 |
|
(b)zurich_city_00_a𝑧𝑢𝑟𝑖𝑐ℎ_𝑐𝑖𝑡𝑦_00_𝑎zurich\_city\_00\_aitalic_z italic_u italic_r italic_i italic_c italic_h _ italic_c italic_i italic_t italic_y _ 00 _ italic_a | 2402.15584v3 | 7 |
|
(c)zurich_city_00_b𝑧𝑢𝑟𝑖𝑐ℎ_𝑐𝑖𝑡𝑦_00_𝑏zurich\_city\_00\_bitalic_z italic_u italic_r italic_i italic_c italic_h _ italic_c italic_i italic_t italic_y _ 00 _ italic_b | 2402.15584v3 | 8 |
|
(d)zurich_city_07_a𝑧𝑢𝑟𝑖𝑐ℎ_𝑐𝑖𝑡𝑦_07_𝑎zurich\_city\_07\_aitalic_z italic_u italic_r italic_i italic_c italic_h _ italic_c italic_i italic_t italic_y _ 07 _ italic_a | 2402.15584v3 | 9 |
|
Figure 1:Temporal coherence of instance representations. Our algorithm encourages instance representations in the BEV space to be coherent across time and viewpoints. | 2402.15583v1 | 1 |
|
Figure 2:Overview of Cohere3D. We construct long-term instance-level temporal correspondence via unlabeled LiDAR point clouds, and design a contrastive learning framework to encourage the long-term coherence of instance representations. After pretraining, the online network is transferred to downstream task-specific models. | 2402.15583v1 | 2 |
|
Figure 3:Construction of long-term temporal correspondence. The first and last scans corresponding to each frame are correlated by clustering. The Hungarian algorithm connects every two neighboring frames to generate long-term correspondence. | 2402.15583v1 | 3 |
|
Figure 6:LiDAR-guided long-term correspondence. It can reliably track each instance over long-term historical frames even if it is captured by different cameras from varying angles. Readers may zoom in to see the projected point clouds clearly. | 2402.15583v1 | 6 |
|
(a)Cohere3D pretraining identifies the boundary of the road to provide a map-compliant plan for turning. | 2402.15583v1 | 7 |
|
(a)Cohere3D pretraining identifies the boundary of the road to provide a map-compliant plan for turning. | 2402.15583v1 | 8 |
|
(b)Cohere3D pretraining detects the car from the opposite direction to avoid potential conflicts during planning. | 2402.15583v1 | 9 |
|
Figure 1:Illustration of Different Types of Supervision in Procedural Learning: Language Instructions, Intermediate Visual Observations, and Task Class. In our project, we have removed all the intermediate supervision and instead rely on captions of visual start and goal as context information. This figure is modified from the work ofWang et al. (2023a). | 2402.15579v1 | 1 |
|
Figure 2:Motivations for our approach. On the left, we observe that when provided with a task name, ChatGPT can generate a detailed plan. However, a proper prompt containing sufficient information is always required for ChatGPT to retrieve knowledge from large language models. In our project, the model is not aware of the task name, which makes the task more challenging and currently unsolved by large language models. On the right, we hypothesize that all the intermediate supervisions, such as language instructions, visual observations, and task class, serve as context information for plan generation. Based on this assumption, we propose a cost-effective approach to obtain context information by generating captions for visual start and goal observations using a pre-trained BLIP modelLi et al. (2022). | 2402.15579v1 | 4 |
|
Figure 3:Overview of our model. Initially, visual observations of the starting and target states (represented by black nodes) are incorporated into the sequence of learned queries (represented by colored nodes). Then stochastic noise is injected into the resulting input sequence. Next, this input is fed into the transformer decoder, which interacts with the global memory to create executable procedural plans. Along with the memory, the context information transformed from the visual start and goal is injected into the transformer decoder. Subsequently, action vectors are generated, and multiple loss functions are employed, to train the model. Specifically, the context information is trained with contrastive learning loss and supervised by captions generated via BLIPLi et al. (2022). Note that this figure is modified fromZhao et al. (2022)and our modifications are highlighted with red boxes. | 2402.15579v1 | 5 |
|
Figure 4:Context augmented transformer decoder. In this architecture, the learned memory bank provides the multi-head cross-attention with K and V. In addition, the learned context feature is concatenated with the input as Q. Note that this figure is modified fromZhao et al. (2022)and our modification is highlighted with the red box. | 2402.15579v1 | 6 |
|
Figure 1.Model Architecture: FC=fully-connected layer, EDL Loss=Evidential Deep Learning Loss Function (Derived using the expected value of the cross-entropy loss over the predicted Beta distribution) | 2402.15572v1 | 1 |
|
Figure 2.Comparison with Original OIA: the color of bounding boxes reflects the rank of importance (red to yellow = most to least important), and the color of actions and explanations reflects the correctness (green=correct, red=incorrect, gray=missing) | 2402.15572v1 | 2 |
|
Figure 3.Comparative Efficacy of Enhanced Models: Bounding box colors indicate importance rank (red to yellow = most to least), while action and explanation colors signify correctness (green=correct, red=incorrect, gray=missing) | 2402.15572v1 | 3 |
|
Figure 4.Sample Failure Cases: the color of bounding boxes reflects the rank of importance (red to yellow = most to least important), and the color of actions and explanations reflects the correctness (green=correct, red=incorrect, gray=missing) | 2402.15572v1 | 4 |
|
Figure 2:Tradeoff between ASR and time for BEAST on Vicuna-7B, by varying our attack parameterk𝑘kitalic_k. We get 98% ASR in 2.65 minutes, while we get 66% ASR in just 10 seconds. | 2402.15570v1 | 2 |
|
Figure 3:Screenshot showing the format and questions asked in human study for evaluating jailbreaking. | 2402.15570v1 | 4 |
|
Figure 4:Screenshot showing the format and questions asked in human study for evaluating hallucination. | 2402.15570v1 | 5 |
|
Figure 5:ASR of BEAST for jailbreaking in the presence of perplexity filter-based defense(Alon & Kamfonas,2023; Jain et al.,2023). We find that there is a trade-off between the readability of the adversarial prompts and the ASR.k=7𝑘7k=7italic_k = 7seems to be the best attack parameter for maintaining good readability and ASR. | 2402.15570v1 | 6 |
|
Figure 1:Example note signal (left) and its corresponding power spectrogram (right) from the NSynth dataset(Engel et al.,2017), which is used to train the proposed self-supervised framework. | 2402.15569v1 | 1 |
|
Figure 2:HCQT power spectrogram (in dB) for an audio signal (top-left), the corresponding first-harmonic channel (bottom-left) and weighted harmonic average (top-right), and an example multi-pitch salience-gram estimate (bottom-right). Self-supervised objectives (ℒharsubscriptℒℎ𝑎𝑟\mathcal{L}_{har}caligraphic_L start_POSTSUBSCRIPT italic_h italic_a italic_r end_POSTSUBSCRIPTandℒsupsubscriptℒ𝑠𝑢𝑝\mathcal{L}_{sup}caligraphic_L start_POSTSUBSCRIPT italic_s italic_u italic_p end_POSTSUBSCRIPT) are employed to encourage the model to concentrate energy around strong F0 candidates. | 2402.15569v1 | 2 |
|
Figure 3:Examples (by row) of sampled Gaussian equalization curves (center) applied to CQT spectrograms (left) to produce equalized spectrograms (right) for the timbre-invariance objective. | 2402.15569v1 | 3 |
|
Figure 4:Examples (by row) of sampled geometric transformations (center) applied to CQT spectrograms (left) to produce transformed spectrograms (right) for the geometric-equivariance objective. | 2402.15569v1 | 4 |
|
Figure 5:CQT spectrogramX1subscript𝑋1X_{1}italic_X start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPTfor track01-AchGottundHerrof Bach10(Duan et al.,2010)along with the multi-pitch salience-gram output generated using our full self-supervised framework (SS-MPE) and the output for the framework with training objective ablations. | 2402.15569v1 | 5 |
|
Figure 6:Positive activation (left), negative activation (center), and standard (right) BCE loss over continuous-valued targets and estimates. | 2402.15569v1 | 6 |
|
Figure 1:Illustration of HILPs.(left) We first train adistance-preservingmappingϕ:𝒮→𝒵:italic-ϕ→𝒮𝒵\phi:{\mathcal{S}}\to{\mathcal{Z}}italic_ϕ : caligraphic_S → caligraphic_Zthat maps temporally similar states to spatially similar latent states (d∗superscript𝑑d^{*}italic_d start_POSTSUPERSCRIPT ∗ end_POSTSUPERSCRIPTdenotes the temporal distance).
(right) We then train a latent-conditioned policyπ(a∣s,z)𝜋conditional𝑎𝑠𝑧\pi(a\mid s,z)italic_π ( italic_a ∣ italic_s , italic_z ), which we call a Hilbert foundation policy, thatspansthat latent space with directional movements. This policy captures diverse long-horizon behaviors from unlabeled data,
which can be directly used to solve a variety of downstream tasks efficiently, even in a zero-shot manner. | 2402.15567v2 | 1 |
|
Figure 2:Diagram of HILPs.(a) We train a Hilbert representationϕ(s)italic-ϕ𝑠\phi(s)italic_ϕ ( italic_s )using a goal-conditioned value learning objective
with the value function parameterized asV(s,g)=−‖ϕ(s)−ϕ(g)‖𝑉𝑠𝑔normitalic-ϕ𝑠italic-ϕ𝑔V(s,g)=-\|\phi(s)-\phi(g)\|italic_V ( italic_s , italic_g ) = - ∥ italic_ϕ ( italic_s ) - italic_ϕ ( italic_g ) ∥.
(b) We train a Hilbert foundation policyπ(a∣s,z)𝜋conditional𝑎𝑠𝑧\pi(a\mid s,z)italic_π ( italic_a ∣ italic_s , italic_z )using the intrinsic reward functionr(s,z,s′)𝑟𝑠𝑧superscript𝑠′r(s,z,s^{\prime})italic_r ( italic_s , italic_z , italic_s start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT )defined as the inner product betweenϕ(s′)−ϕ(s)italic-ϕsuperscript𝑠′italic-ϕ𝑠\phi(s^{\prime})-\phi(s)italic_ϕ ( italic_s start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ) - italic_ϕ ( italic_s )and a randomly sampled unit vectorz𝑧zitalic_z. | 2402.15567v2 | 2 |
|
Figure 3:Test-time midpoint planning.In the presence of embedding errors,
the direction toward the midpoint subgoalw∗superscript𝑤w^{*}italic_w start_POSTSUPERSCRIPT ∗ end_POSTSUPERSCRIPTcan be more accurate than the direction toward the goalg𝑔gitalic_g. | 2402.15567v2 | 3 |
|
Figure 5:Zero-shot RL performance.HILP achieves the best zero-shot RL performance in the ExORL benchmark,
outperforming previous state-of-the-art approaches.
The overall results are aggregated over4444environments,4444tasks,4444datasets, and4444seeds (i.e.,256256256256values in total). | 2402.15567v2 | 5 |
|
Figure 6:Pixel-based zero-shot RL performance.HILP exhibits the best performance in thepixel-basedExORL benchmark as well,
outperforming the two most performant baselines among FB and SF methods.
The results are aggregated over4444environments,4444tasks,4444datasets, and4444seeds (i.e.,256256256256values in total). | 2402.15567v2 | 8 |
|
Figure 7:Planning performance in AntMaze (𝟖8\mathbf{8}bold_8seeds).Test-time midpoint planning (Section5.3) improves performance
as the number of recursive refinements increases. | 2402.15567v2 | 9 |
|
Figure 8:Ablation study of latent dimensions for zero-shot RL.In the ExORL benchmark,D=50𝐷50D=50italic_D = 50generally leads to the best performance across the environments.
The results are aggregated over4444tasks,4444datasets, and4444seeds (i.e.,64646464values in total). | 2402.15567v2 | 10 |
|
Figure 9:Ablation study of latent dimensions for offline goal-conditioned RL (𝟒4\mathbf{4}bold_4seeds).In D4RL tasks,D=32𝐷32D=32italic_D = 32generally leads to the best performance across the environments. | 2402.15567v2 | 11 |
|
Figure 10:Embedding error analysis (𝟒4\mathbf{4}bold_4seeds).We show the relationship between performance and Hilbert embedding errors on the antmaze-large-diverse task.
In general, lower embedding errors lead to better goal-reaching performance. | 2402.15567v2 | 12 |
|
Figure 11:Visualization of Hilbert representations.We visualize Hilbert representations learned on the antmaze-large-diverse dataset.
Since Hilbert representations are learned to capture the temporal structure of the MDP,
they focus on the global layout of the maze even when we use a two-dimensional latent space (D=2𝐷2D=2italic_D = 2),
and accurately capture the maze layout with a32323232-dimensional latent space (D=32𝐷32D=32italic_D = 32). | 2402.15567v2 | 13 |
|
Figure 12:Per-environment performances on the state-based ExORL benchmark.The results are aggregated over4444tasks,4444datasets, and4444seeds (i.e.,64646464values in total). | 2402.15567v2 | 14 |
|
Figure 13:Per-dataset performances on the state-based ExORL benchmark.The results are aggregated over4444environments,4444tasks, and4444seeds (i.e.,64646464values in total). | 2402.15567v2 | 15 |
|
Figure 14:Per-task performances on the state-based ExORL benchmark.The results are aggregated over4444datasets and4444seeds (i.e.,16161616values in total). | 2402.15567v2 | 16 |
|
Figure 15:Per-environment performances on the pixel-based ExORL benchmark.The results are aggregated over4444tasks,4444datasets, and4444seeds (i.e.,64646464values in total). | 2402.15567v2 | 17 |
|
Figure 16:Per-dataset performances on the pixel-based ExORL benchmark.The results are aggregated over4444environments,4444tasks, and4444seeds (i.e.,64646464values in total). | 2402.15567v2 | 18 |
|
Figure 17:Per-task performances on the pixel-based ExORL benchmark.The results are aggregated over4444datasets and4444seeds (i.e.,16161616values in total). | 2402.15567v2 | 19 |
|
Figure 3.Averaged gate reduction per circuit size for (a) AltGraph GRU, (b) AltGraph GCN,(c) AltGraph DeepGMG. | 2403.12979v2 | 3 |
|
Figure 4.Averaged depth reduction per circuit size for (a) AltGraph GRU, (b) AltGraph GCN, (c) AltGraph DeepGMG. | 2403.12979v2 | 4 |
|
Figure 2:Attention scores of a single query pose (current frame) as a function of the pose attended to (x-axis) in a diffusion-based motion generation model with a sinusoidal absolute positional encoding. Curves show the scores at each denoising step. We observe that, whereas early steps show strong global dependencies (blue), later denoising stages exhibit a clearly local behavior (red). | 2402.15509v1 | 2 |
|
Figure 3:Pose-centric cross-attention.Our attention minimizes the entanglement between the control signal (e.g., text, objects) and the noisy motion by feeding the former only to the query. Consequently, our model denoises each frame’s noisy pose only leveraging its own condition, and the neighboring noisy poses. | 2402.15509v1 | 3 |
|
Figure 4:Transitions smoothness.Average maximum jerk over joints at each frame of the transitions for both motion composition (left) and extrapolation (right) tasks. While other methods show severe smoothness artifacts in the beginning and end of their transition refinement processes,FlowMDM’s jerk curve has the shortest peak for composition, and an absence of peaks for extrapolation. | 2402.15509v1 | 4 |
|
Figure 5:BPE trade-offs.Increasing the number of APE steps undergone during BPE sampling improves the correspondence between motion and textual description (R-prec), but reduces the transition realism and smoothness (FID and AUJ). The best balance is reached around 10% of APE denoising steps. | 2402.15509v1 | 5 |
|
Figure A:Diffusion noise schedules.The cosine noise schedule destroys the motion signal slower and in a more evenly distributed way than the linear schedule. As a result,FlowMDMis able to exploit better the low-to-high frequencies decomposition along the denoising chain and generate better subsequences and transitions. The faster motion destruction in the linear schedule translates to needing more APE steps to reconstruct global dependencies inside subsequences (black arrows↔normal-↔\leftrightarrow↔). | 2402.15509v1 | 7 |
|
Figure B:Classifier-free guidance.In line with prior works, we also observe an accuracy improvement (R-prec) when increasing the strength (i.e.,weight) of the classifier-free guidance (CFG). However, above certain values, the performance degrades, especially in terms of smoothness (AUJ). This is caused by the misalignment of CFG directions on each side of the transition. | 2402.15509v1 | 8 |
|
Figure C:Qualitative examples (Babel).A-F feature six human motion compositions, and G-H two human motion extrapolations. According to the scenarios defined inSec. B, A, B, C belong to in-distribution combinations, and D, E, F to out-of-distribution combinations. Videos of all samples are also included as part of this supplementary material. Solid curves match the trajectories of the global position (blue) and left/right hands (purple/green). Darker colors indicate instantaneous jerk deviations from the median value, saturating at twice the jerk’s standard deviation in the dataset (black segments). Abrupt transitions manifest as black segments amidst lighter ones. | 2402.15509v1 | 9 |
|
Figure D:Qualitative examples (HumanML3D).A-F feature six human motion compositions, and G-H two human motion extrapolations. According to the scenarios defined inSec. B, A, B, C are samples from the short, medium, and long scenarios, respectively, and D, E, F from the mixed scenario. Videos of all samples are also included as part of this supplementary material. | 2402.15509v1 | 10 |
|
Figure 5:Deformation with depth.Change of average eccentricity(Xu et al.,2021)of the input space neighborhoods𝑽xsubscript𝑽𝑥{\bm{V}}_{x}bold_italic_V start_POSTSUBSCRIPT italic_x end_POSTSUBSCRIPTby different layers of a CNN trained on the CIFAR10 dataset, for different radiusr𝑟ritalic_r.
We see that, for larger radius, the deformation increases with depth almost exponentially. Forr≤0.014𝑟0.014r\leq 0.014italic_r ≤ 0.014deformation is low, indicating that smaller radius neighborhoods are reliable for LC computation on deeper networks. Values are averaged over neighborhoods sampled for1000100010001000training points from CIFAR10. For ResNet18, seeFigure23. | 2402.15555v2 | 5 |
|
Figure 6:Grokking across datasets and architectures.From left to right, examples of delayed robustness emerging late in training for a CNN trained on CIFAR10, CNN trained on CIFAR100, and ResNet18 trained on the Imagenette222github.com/fastai/imagenettedataset. Clear double descent behavior visible in the local complexity of CNN with CIFAR10 and CIFAR100. The ResNet18 trained with Imagenette obtains a very high local complexity during the ascent phase of training. To compute local complexity we consider25252525dimensional neighborhoods centered on1024102410241024train, test or random samples. We user=0.005𝑟0.005r=0.005italic_r = 0.005for CNN andr=10−4𝑟superscript104r=10^{-4}italic_r = 10 start_POSTSUPERSCRIPT - 4 end_POSTSUPERSCRIPTfor ResNet18. | 2402.15555v2 | 6 |
|
Figure 7:Grokking visualized.We induce grokking by randomly initializing a4444depth200200200200width ReLU MLP and scaling the initialized parameters by eight following(Liu et al.,2022). In the leftmost figure, we can see that the grokking is visible for both the test samples as well as adversarial examples generated using the test set. We see that the network robustness, periodically increases. By visualization the partition and curvature of the function across a 2D slice of the input space(Humayun et al.,2023a), we see that the network periodically increases the concentration of non-linearity around its decision boundary, making the boundary sharper at each robustness peak. This occurs even when the network doesn’t undergo delayed generalization (Figure2). As the local complexity around the decision boundary increases, the local complexity around data points farther from the decision boundary decreases (Figure26). | 2402.15555v2 | 7 |
|
Figure 8:Region migration in modular addition.By measuring the local complexity for the GeLU activated fully connected layers of a Transformer architecture, we see that here as well, region migration occurs during grokking. | 2402.15555v2 | 8 |
|
Figure 9:Delayed robustness in LLMs.Grokking observed in a GPT architecture with 12 heads and 12 layers trained on next character prediction using the Shakespeare Text Dataset. We see that the second local complexity descent starts prior to the test acc. peak, and descent continues while the network groksϵ=0.03italic-ϵ0.03{\epsilon}=0.03italic_ϵ = 0.03ℓ∞subscriptℓ\ell_{\infty}roman_ℓ start_POSTSUBSCRIPT ∞ end_POSTSUBSCRIPT-PGD adversarial examples, generated in the token embedding space. Approximate input space partition visualized inFigure19andFigure20. | 2402.15555v2 | 9 |
|
Figure 10:Local complexity across depths.From left to right, accuracy, local complexity around training and local complexity around test data points, for an MLP trained on MNIST with width200200200200and varying depth.As depth is increased the max LC during ascent phase becomes larger. We can also see a distinct second peak right before the descent phase. | 2402.15555v2 | 10 |
|
Figure 11:Token embedding space partition formed by the GeLU activated MLP of Block 12 of the GPT model mentioned inFigure9. The partition is approximated by computing LC on a512×512512512512\times 512512 × 512grid on a 2D subspace in the token embedding space. Note that after grokking (right), the three random samples used to determine the 2D subspace, has visibly lower local complexity in its immediate neighborhood. | 2402.15555v2 | 11 |
|
Figure 12:Batch-norm removes grokking.Training a CNN with an identical setting as inFootnote2-left, except the CNN now has Batch Normalization layers after every convolution. With the presence of batchnorm, the LC values increase, the initial descent gets removed and most importantly, grokking does not occur for adversarial samples. | 2402.15555v2 | 12 |
|
Figure 13:Memorization requirement delays grok.When training an MLP on varying number of randomly labeled MNIST samples, we see that with increase in the number of samples, the local complexity dynamics get delayed, especially the ascent phase gets elongated. This shows that with increased demand for memorization the network takes longer to complete ascent and later undergo region migration. | 2402.15555v2 | 13 |
|
Figure 14:Increasing width hastens region migration.LC dynamics while training an MLP withvarying widthon MNIST. For the peak LC achieved around training points during the ascent phase, we see an initial increase and then decrease as the network gets overparameterized. For test and random samples, we see the LC during ascent phase saturating as we increase width. | 2402.15555v2 | 14 |
|
Figure 15:LC for aP𝑃Pitalic_Pdimensional neighborhood with radiusr𝑟ritalic_rwhile being shifted from the origin[0]dsuperscriptdelimited-[]0𝑑[0]^{d}[ 0 ] start_POSTSUPERSCRIPT italic_d end_POSTSUPERSCRIPTto vector[10]dsuperscriptdelimited-[]10𝑑[10]^{d}[ 10 ] start_POSTSUPERSCRIPT italic_d end_POSTSUPERSCRIPT. Inleft, we varyP𝑃Pitalic_Pwith fixedr=5𝑟5r=5italic_r = 5while onrightwe varyr𝑟ritalic_rfor fixedP=20𝑃20P=20italic_P = 20. We see that for all the settings, shifting away from the origin reduces LC. The increase of LC with the neighborhood dimensionalityP𝑃Pitalic_Pgets saturated as we increaseP𝑃Pitalic_P, showing that lower dimensional neighborhoods can be good enough for approximating LC. Increasingr𝑟ritalic_ron the other hand, increases LC and reduces LC variations between shifts, since the neighborhood becomes larger and LC becomes less local. | 2402.15555v2 | 15 |
|
Figure 16:Training accuracy and robust accuracy for the networks trained on randomly labeled MNIST samples presented inFigure13. | 2402.15555v2 | 16 |
|
Figure 17:Increasing the batch-size expedites grokking. This indicates that reduced SGD noise allows region migration to occur earlier in training. | 2402.15555v2 | 17 |
|
Figure 18:Grokking stronger adversarial attacks.We see that during delayed generalization, the robustness to Auto-Attack(Croce & Hein,2020)also increases. This shows the universality of delayed robustness. | 2402.15555v2 | 18 |
|
Figure 19:Token embedding space LC on a 2D subspace intersecting three random training points. We visualize the LC layerwise for the GeLU activated MLP layers inside each of the 12 blocks of the LLM for which we present training dynamics inFigure9. The LC is computed after 197 optimization steps, during the peak ascent. We see that LC on this subspace is very high especially close to the data points. LC values are clamped to a maximum of 150. | 2402.15555v2 | 19 |
|
Figure 20:Token embedding space LC on a 2D subspace intersecting three random training points. We visualize the LC layerwise for the GeLU activated MLP layers inside each of the 12 blocks of the LLM for which we present training dynamics inFigure9. The LC is computed after 372759 optimization steps, therefore during the second LC descent. We see that LC on this subspace is concentrated away from the training points, especially for the deeper layers. This indicates that region migration occurs in LLMs as well, leading to delayed robustness. LC values are clamped to a maximum of 150. | 2402.15555v2 | 20 |
|
Figure 21:Change of avg. eccentricity and diameter(Xu et al.,2021)of the input space neighborhood by different layers of a CNN trained on the CIFAR10 dataset. For different sampling radiusr𝑟ritalic_rof the sampled input space neighborhoodV𝑉Vitalic_V, the change of eccentricity and diameter denotes how much deformation the neighborhood undergoes between layers. Here, layer00corresponds to the input space neighborhood. Numbers are averaged over neighborhoods sampled for1000100010001000training points from CIFAR10.
For larger radius the deformation increases with depth exponentially. Forr≤0.014𝑟0.014r\leq 0.014italic_r ≤ 0.014deformation is lower, indicating that smaller radius neighborhoods are reliable for LC computation on deeper networks. Confidence interval shown in red, is almost imperceptible. | 2402.15555v2 | 21 |
|
Figure 22:Training a ResNet18 with batchnorm on Imagenet Full. LC is computed only on test points using 1000 test set samples. Computing LC 1000 samples takes approx. 28s on an RTX 8000. | 2402.15555v2 | 22 |
|
Figure 23:Change of avg. eccentricity and diameter(Xu et al.,2021)of the input space neighborhood by different layers of a ResNet18 trained on the CIFAR10 dataset, similar to the setting of Fig.21. Resnet deforms the input neighborhood by reducing the avg. eccentricity and diameter of the neighborhood graphs. Forr≤0.014𝑟0.014r\leq 0.014italic_r ≤ 0.014deformation is lower, indicating that smaller radius neighborhoods are reliable for LC computation on deeper networks. | 2402.15555v2 | 23 |
|
Figure 24:Layerwise visualization of the input space partition for a 2D domain passing through a training set triad, after robust partition formation. The partition is visualized for an MLP with depth 6 and width 200, trained on1000100010001000samples from MNIST, similar to the setting described inFigure2. We see that deeper layer neurons partake more in the formation of the robust partition, compared to shallower layers. This is due to the fact that deeper layer neurons can be more localized in the input space due to the non-linearity induced by preceding layers. | 2402.15555v2 | 24 |
|
Figure 25:Input space partition computed analytically via SplineCam(Humayun et al.,2023b)for the 2D toy setting presented inFigure3(left). Regions are colored by white and knots are colored by red. The partition is computed for the input space domain[−10,10]2superscript10102[-10,10]^{2}[ - 10 , 10 ] start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT, induced by an MLP of depth5555and width30303030. We take a meshgrid of300×300300300300\times 300300 × 300points over the input domain, and measure the local complexity at each point with radius,r∈{0.05,0.1,0.5,1}𝑟0.050.10.51r\in\{0.05,0.1,0.5,1\}italic_r ∈ { 0.05 , 0.1 , 0.5 , 1 }(rest). We see that our proposed method can locate the non-linearities for smallr𝑟ritalic_r. Asr𝑟ritalic_ris increased our method provides a coarser estimate of the local density of non-linearities, i.e., number of non-linearities intersecting the a fixed volume defined by the local neighborhood. | 2402.15555v2 | 25 |
|
Figure 26:Partition visualization for 2D domains localized around the decision boundary (top) and away from the decision boundary (bottom) for the grokking setup presented inFigure7. All the plots are show for the optimization step95381953819538195381. Number of regions in the partition for top-right, top-left, bottom-right, and bottom-left are 123156, 88362, 33273, and 32018 respectively. The domain used for all of the plots has the same area/volume. Therefore, close to the decision boundary, the region density is much higher compared to away from the decision boundary. This is evidence of region migration happening during the latter phases of training. | 2402.15555v2 | 26 |
|
Figure 27:MLP with width 200 and varying depth being trained on1000100010001000samples from MNIST. Increasing the depth of the network decreases the sharpness of the LC peak during ascend phase. Deeper networks also tend to have a sharper decline in the training LC during region migration. | 2402.15555v2 | 27 |
|
Figure 28:For an MLP with depth 4 and width 200, we train with varying training set sizes and evaluate the adversarial performance after each training iteration. We see that with increasing dataset size, the network groks earlier in time, as can be visible in the adversarial grokking curves for all the different epsilon values. | 2402.15555v2 | 28 |
|
Figure 29:Training a 200 width MLP on MNIST with initialization scaling of 8 and varying depths. Along the row, we consider larger and larger radius neighborhoods for local complexity approximation. | 2402.15555v2 | 29 |
|
Figure 31:Increasing the volume of randomly labeled training data.Continued fromFigure13. Increasing the number of randomly labeled training samples delays the ascent phase of the LC training dynamics for both training and test samples. For random samples the behavior is not affected as much. | 2402.15555v2 | 31 |
|
Figure 32:Increasing training data size expedites region migration.Local complexity dynamics training an MLP on MNIST with weight decay. Robustness plots presented inFigure28. | 2402.15555v2 | 32 |
|
Figure 34:LC dynamics for a GeLU-MLP with width200200200200and depth{3,4,5}345\{3,4,5\}{ 3 , 4 , 5 }presented from left to right. LC is calculated at1000100010001000training points and10000100001000010000test and random points during training on MNIST. | 2402.15555v2 | 34 |
|
Figure 35:Comparing the local complexity measured in terms of the number of linear regions computed exactly by SplineCAM(Humayun et al.,2023a)and number of hyperplane cuts by our proposed method. Both methods exhibit the double descent behavior. | 2402.15555v2 | 35 |
|
Figure 2:Effectiveness of vanilla weak-to-strong generalization.
The performance gap recovery (PGR) is notable when the performance of the supervisor is close to the ceiling performance of the strong student (0.74), but limited when the supervisor lags behind. | 2402.15505v1 | 2 |
Subsets and Splits