image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
(b)Embeddings ofS1superscript𝑆1S^{1}italic_S start_POSTSUPERSCRIPT 1 end_POSTSUPERSCRIPTintoℝ2superscriptℝ2\mathbb{R}^{2}blackboard_R start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT,p𝑝pitalic_pvarying,L𝐿Litalic_Lmeasured with Euclidean distance | 2403.07929v1 | 20 |
|
Figure 3:Confusion matrices for five different classification models. (a) Random Forest model. (b) SVM (LibSVM) model. (c) AdaBoost model. (d) Naive Bayes model. (e) Bagging with Decision Trees model. | 2403.01022v1 | 3 |
|
Figure 5:Box Plot of Classification Model Cross-Validation Results. The distribution of accuracy scores from five-fold cross-validation for five different classification models is shown in this figure: AdaBoost, Random Forest, SVM (LibSVM), Naive Bayes, and Decision Tree-Based Bagging. | 2403.01022v1 | 15 |
|
Figure 1:Pessimism adjustment can yield performance benefits exceeding those of increased replay ratio and full-parameter resets. The pessimistic algorithm dominates Humanoid, whereas the optimistic algorithm dominates Hopper. 10 seeds and 95% CI. | 2403.01014v1 | 1 |
|
Figure 3:High level overview of the proposed approach. After environment step, the transition is stored in either the training buffer (used for updating actor-critic modules) or the validation buffer (used for updating pessimism module). The pessimism is updated via a ”reverse” TD loss, optimisation of which on the training buffer would be prone to overfitting. | 2403.01014v1 | 8 |
|
Figure 7:VPL exhibits substantially less sensitivity to the learning rate of the pessimism module. 4 tasks, 10 seeds per task. | 2403.01014v1 | 33 |
|
Figure 10:The critic disagreement of SAC algorithm does not completely diminish in the considered training regime of 1 million environment steps, making adjustments to the pessimism a viable strategy. 10 seeds per task. | 2403.01014v1 | 41 |
|
Figure 1:Top: Training curves of model-free PO for(P1)-(P3)averaged over6666random seeds with shaded regions denoting standard deviation. We normalize the vertical axis with respect to the cost of the LQT controller based on the DMDc model. For(P1), the training process of PO (without a warm start) instantly destabilizes withη=10−4𝜂superscript104\eta=10^{-4}italic_η = 10 start_POSTSUPERSCRIPT - 4 end_POSTSUPERSCRIPT. Hence, we selectη=10−5𝜂superscript105\eta=10^{-5}italic_η = 10 start_POSTSUPERSCRIPT - 5 end_POSTSUPERSCRIPT, the largest learning rate under which PO can consistently decrease the cost.Bottom: The tracking costs of(P1)-(P3)with no control, LQT, LQT-PO after 40 iterations, and pure PO after 40 iterations. The costs are averaged over 10 trajectories with randomly sampled initial fieldsu(x,t=0)𝑢𝑥𝑡0u(x,t=0)italic_u ( italic_x , italic_t = 0 ). The shaded region denotes the standard deviation. | 2403.01005v1 | 3 |
|
Figure 2:(P1):Comparing model-based, PO with warm start, and pure PO control strategies with40404040iterations of training budget for the latter two. | 2403.01005v1 | 4 |
|
Figure 3:(P2):Comparing model-based, PO with warm start, and pure PO control strategies with40404040iterations of training budget for the latter two. | 2403.01005v1 | 5 |
|
Figure 4:(P3):Comparing model-based, PO with warm start, and pure PO control strategies with40404040iterations of training budget for the latter two. | 2403.01005v1 | 6 |
|
Figure 1:Overview of Attribute Structuring LLM evaluation for clinical summarization.
AS uses a human-given ontology to extract attributes from summaries and then to individually score them. | 2403.01002v2 | 1 |
|
Figure 2:Distribution of scores across different attributes.
Attribute Structuring enables auditing which aspects of discharge summaries are generally accurate (e.g.Main diagnosis) and which are often inaccurate (e.g.Lab test). | 2403.01002v2 | 2 |
|
Figure 3:Example interpretation of document similarity.
Scores for each element of the ontology can be analyzed and grounded in quotes from the source summaries for better understanding. | 2403.01002v2 | 3 |
|
Figure 1:Three-dimensional evaluation on methods that scale to ImageNet-1K.Left:Recovery accuracy vs. storage trade-off comparison for our (D3) and other methods on resized (64×64×36464364\times 64\times 364 × 64 × 3) ImageNet-1K. Our method achieves SOTA performance at small memory cost regime.Right:Accuracy vs. downstream task training cost on resized ImagNet-1K. | 2403.00999v1 | 1 |
|
Figure 2:Illustration of Federated Distillation and Distributional RepresentationWe decompose large datasets into subtasks and distill each subset into distributions using locally trained experts. Distributions distilled on subtasks generalize well to the full task. | 2403.00999v1 | 2 |
|
Figure 3:Visualization of distilled mean and variations for four classes from ImageNet-1Kfirst column:typical images for each class generated by passing the mean to the decoder to generate the sample.second column onwards:variations by sampling from the corresponding latent distribution | 2403.00999v1 | 3 |
|
Figure 4:Visualization of the latent Gaussian space by interpolating priors for four classes from ImageNet-1KThe four corners are the mean for each class. We linearly sample the Gaussian space between classes and pass into the decoder to generate interpolated images. | 2403.00999v1 | 4 |
|
Figure 5:Ablation study on dual training objective and distributional representationBoth distributional representation and dual training objective are essential for the performance of our method. | 2403.00999v1 | 5 |
|
Figure 6:Visualization of distilled samples from five classes using different training objectives and fixed or distributional representation on TinyImageNetfor distributional outcomes we visualize the mean. | 2403.00999v1 | 6 |
|
Figure 7:ImageNet-1K recovery accuracy using distilled labels or hard labelsOur method still performs well even without using distilled labels. | 2403.00999v1 | 7 |
|
Figure 9:Samples from Distilled Distribution on ImageNet-1K under 100MB storage cost. 2 latent priors per class visualized | 2403.00999v1 | 9 |
|
Figure 10:Samples from Distilled Distribution on ImageNet-1K under 100MB storage cost (continued). 2 latent priors per class visualized | 2403.00999v1 | 10 |
|
Figure 1:Left:example of stimulus material (irony condition) for different methods. Sequences for which scores are retrieved are underlined. The trigger sentence is in boldface.T𝑇Titalic_T,Cisubscript𝐶𝑖C_{i}italic_C start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPTandQ𝑄Qitalic_Qare included in all methods’ input promptsIisubscript𝐼𝑖I_{i}italic_I start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT. Colors indicate the text which is additionally used for the respective method.Right:set of all relevant scores. | 2403.00998v1 | 1 |
|
Figure 2:Task accuracy for different methods and models, separately for different conditions.
Bars show accuracy scores, which are averages over different versions of each method (e.g., random seeds or different scores).
Horizontal bar shows accuracy expected from random guessing in each condition. | 2403.00998v1 | 2 |
|
Figure 3:Results of methods comparison, based on accuracy (proportion of target choices) and goodness-of-fit to the human data. Different scores are displayed on the x-axis.
For ease of visual comparison, goodness-of-fit is shown as difference in log-likelihood of the human data under the LLM predictor, compared against one of the worst performing models (higher is better). | 2403.00998v1 | 3 |
|
Figure 1:Overview of our proposed online learning system, SELFI. Our method fine-tunes a pre-trained control policy trained with model-based objective by incorporating this objective into a Q-value function to maximize during online model-free RL. | 2403.00991v2 | 1 |
|
Figure 2:SELFI architecture overview.Before online learning, we train the encoder and the actor by maximizing the differentiable model-based objective. In the online phase, we combine the offline objective with the learnedQ𝑄Qitalic_Q-value from model-free RL to fine-tune the actor. | 2403.00991v2 | 2 |
|
Figure 4:Visualization of the robot behavior when interacting with the pedestrian.Robots and pedestrians with the same color indicate the same time. The black dotted lines indicate the time of closest proximity. | 2403.00991v2 | 4 |
|
Figure 6:The number of intervention on online learning in three different environments.The lines indicate the mean and the areas indicate the standard deviation. The horizontal axis indicates the number of times the robot laps the loop reference. | 2403.00991v2 | 6 |
|
Figure 9:The number of intervention on online learning in Environment 1.Ours + prior data indicates our method leveraging large dataset in same environment. Residual RL†uses the SACSoN policy asπbasesubscript𝜋base\pi_{\text{base}}italic_π start_POSTSUBSCRIPT base end_POSTSUBSCRIPT. The horizontal axis indicates the number of times the robot laps the loop. | 2403.00991v2 | 9 |
|
Figure 10:Overview of robot’s position estimation with AR markers.We update the correlation matrixTicsubscriptsuperscript𝑇𝑐𝑖T^{c}_{i}italic_T start_POSTSUPERSCRIPT italic_c end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPTwhen the robot detects an AR marker, and we apply it to suppress the error from the noisy visual odometryTtvosubscriptsuperscript𝑇𝑣𝑜𝑡T^{vo}_{t}italic_T start_POSTSUPERSCRIPT italic_v italic_o end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPTuntil the robot detects the next AR marker. | 2403.00991v2 | 10 |
|
Figure 1:A comparison overview of the four formulations we study in this paper: NLI, Pairwise, MRC, and Timeline. A document with event annotation is converted into a different number of prompts following each formulation, and the predictions are interpreted as pairwise temporal orders in each formulation’s manner to form a timeline. | 2403.00990v1 | 1 |
|
Figure 2:Formulation comparison result with TimeSET. In each boxplot, one data point represents a combination of a prompt template and the number of demonstrations. | 2403.00990v1 | 2 |
|
Figure 3:Pseudo-perplexity scores of BERTs, trained on the masked language modeling task, combined using our method. Curves differ by which components they merge. Results across 10 merges are shown with standard error regions shaded around each curve. Each additional merged component leads to further barrier reduction. | 2403.00986v3 | 3 |
|
Figure 4:Average feature correlations between layers from different MultiBERTs. We report correlations for both Feed-Forward and MHA features. Both components see much higher average correlation after applying their respective component permutations. Values are averaged over 10 merges, with standard error regions shaded. | 2403.00986v3 | 4 |
|
Figure 5:Visualization of correlation matrices between features before and after permuting. These features are from the seventh multi-headed attention layer from 2 different MultiBERTs models. On the left, 12 attention head boundaries are clearly visible, and highly correlated regions do not necessarily correspond to the same attention head indices, supporting our two-stage permutation method. On the right, the two-stage permutation method outcome can be seen via the dark diagonal line, and its surrounding block diagonal pattern. | 2403.00986v3 | 5 |
|
Figure 6:Loss barriers on the masked language modeling task with different data amounts used for computing features. Standard error regions are shaded. The amount of data does not have a strong directional relationship with the size of the loss barrier in this setting. | 2403.00986v3 | 6 |
|
Figure 1:Given a collection of documents,LocalRQAprovides tools to generate RQA data, to train and test open-source models, and to deploy the RQA system for human evaluation or as an interactive chatbot. | 2403.00982v1 | 1 |
|
Figure 2:An overview of theLocalRQAtoolkit, which supports the entire pipeline of developing an RQA system: from data processing to training, testing, and serving an RQA system. Different from many existing toolkits, we feature a wide selection of training, testing, and serving methods curated from the latest RQA research. | 2403.00982v1 | 2 |
|
Figure A1:Researchers can launch a human evaluation page usingLocalRQAin a single command line. Given a prediction file (seeSection3.5),LocalRQAlaunches a web server that allows other users to evaluate the quality of pre-generated responses. Evaluation results are automatically saved for researchers to conduct further analysis. | 2403.00982v1 | 4 |
|
Figure A2:Researchers can launch an interactive chat page withLocalRQAusing three commands.LocalRQAuses a model controller back-endZheng et al. (2023a)to handle load-balancing. Chat histories are automatically saved for researchers to conduct further analysis or model training. | 2403.00982v1 | 5 |
|
Figure 2:Mean Ranks of Counterfactual-Free (black) and Counterfactual-Guided (grey) Semi-Factual Methods | 2403.00980v2 | 4 |
|
Figure 4:Errors (solid lines) and1σ1𝜎1\sigma1 italic_σbounds (dashed lines) of the spatial-temporal calibration parameters.x𝑥xitalic_x-axis represents time in seconds. Left to right corresponds to Case1 to Case5 inSec.6.1. The estimation error of the rotation and temporal calibration parameters perfectly approach to zero for any cases. While the convergence results of the translation calibration parameter are varied from case to case. | 2403.00976v1 | 16 |
|
Figure 5:imu1𝑖𝑚𝑢1imu1italic_i italic_m italic_u 1is used. GT: groundtruth trajectory output from motion capture system. PnP: camera trajectory output from PnP algorithm. Ours: refined camera trajectoryTCiW,i=1⋯Nsubscriptsuperscript𝑇𝑊subscript𝐶𝑖𝑖1⋯𝑁{}_{{C_{i}}}^{W}T,{\rm{}}i=1\cdots Nstart_FLOATSUBSCRIPT italic_C start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT end_FLOATSUBSCRIPT start_POSTSUPERSCRIPT italic_W end_POSTSUPERSCRIPT italic_T , italic_i = 1 ⋯ italic_N. | 2403.00976v1 | 22 |
|
Figure 7:Groundtruth (solid lines) and estimation (dashed lines) of the time-varying change of the spatial-temporal parameters. | 2403.00976v1 | 24 |
|
Figure 8:Iterative process of calibrating left camera intrinsic from scratch.x𝑥xitalic_x-axis represents iteration steps. | 2403.00976v1 | 25 |
|
Figure 9:Iterative process of calibrating right camera intrinsic from scratch.x𝑥xitalic_x-axis represents iteration steps. | 2403.00976v1 | 26 |
|
Figure 2:Distributions of z-scores across 350 DNNs pruned to 98% sparsity. Top axis shows the motif, left axis shows the standard deviations from the mean (or z-score), and the right axis shows the cumulative percentage. | 2403.00974v1 | 2 |
|
Figure 3:Z-score distributions across sparsity levels. Each panel shows how the z-score of the pictured motif changes throughout the pruning process. The bottom right panel shows the test MSE across all 350 networks at increasing levels of sparsity. | 2403.00974v1 | 3 |
|
Figure 4:Left: Example of two-layer sparse network with inputs,x→→𝑥\vec{x}over→ start_ARG italic_x end_ARG, and outputs,y→→𝑦\vec{y}over→ start_ARG italic_y end_ARG. Center: Forward pass computation where𝐖𝐖\mathbf{W}bold_Wrepresents the weight matrix𝐌𝐌\mathbf{M}bold_Mrepresents the mask matrix, andσ𝜎\sigmaitalic_σrepresents the nonlinear activation function (bias is excluded for simplicity). Right: Example of weight matrix𝐖𝐖\mathbf{W}bold_Wand mask matrix𝐌𝐌\mathbf{M}bold_M. | 2403.00974v1 | 4 |
|
Figure 5:Left: Example of sparse feed-forward network with inputs,x→→𝑥\vec{x}over→ start_ARG italic_x end_ARG, outputs,y→→𝑦\vec{y}over→ start_ARG italic_y end_ARG, and one hidden layer,h→1subscript→ℎ1\vec{h}_{1}over→ start_ARG italic_h end_ARG start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT. Middle: Same network with second-order chain sub-graphs highlighted. Right: Masks representing the connectivity of the network between the layers (e.g.,𝐌𝐱,𝐡𝟏subscript𝐌𝐱subscript𝐡1\mathbf{M_{x,h_{1}}}bold_M start_POSTSUBSCRIPT bold_x , bold_h start_POSTSUBSCRIPT bold_1 end_POSTSUBSCRIPT end_POSTSUBSCRIPTfor the weights between layersx→→𝑥\vec{x}over→ start_ARG italic_x end_ARGandh→1subscript→ℎ1\vec{h}_{1}over→ start_ARG italic_h end_ARG start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT). | 2403.00974v1 | 5 |
|
Figure 6:Example of sparse, feed-forward network converted to list of masks, ormask listas referred to in algorithms. | 2403.00974v1 | 6 |
|
Figure 7:Total sub-graph count in 350 sparse networks across sparsity levels. Each panel gives results for each sub-graph type. Used in the z-score calculation to produce results in Figure3. | 2403.00974v1 | 7 |
|
Figure 8:Average sub-graph count across 1000 random networks generated for each of the 350 sparse networks across sparsity levels. Each panel gives results for each sub-graph type. Used in the z-score calculation to produce results in Figure3. | 2403.00974v1 | 8 |
|
Figure 9:Standard deviation of the sub-graph count across 1000 random networks generated for each of the 350 sparse networks across sparsity levels. Each panel gives results for each sub-graph type. Used in the z-score calculation to produce results in Figure3. | 2403.00974v1 | 9 |
|
Figure 1:Pipeline architecture depicting data augmentation techniques and weighted ensemble of strategies | 2403.00964v1 | 1 |
|
Figure 5:Pseudocode of in-batch transformation for T2V in a PyTorch-like style. Step 1 replaces pairwise comparison with matrix manipulation, while Step 2 showcases on-the-fly transformations for mini-batch implemented through thetranforms.Composeformulae-sequencetranformsCompose\mathrm{tranforms.Compose}roman_tranforms . roman_Composemodule in PyTorch. | 2403.00963v1 | 4 |
|
(b)T2T: binary encode. A pseudo node G is added to make the tree complete and infilled with0.50.50.50.5by default. | 2403.00963v1 | 11 |
|
Figure 4:This flowchart shows the relationships between the modules and their topic (general data science or machine learning). The arrows depict suggested prerequisites for the relevant data science and machine learning topics, starting from the most basic module DSECOP 101: Introduction to Data Science Libraries. Introductory, intermediate, and advanced-level modules are denoted by their borders (solid, dashed, and dotted) and also by their module number (1XX, 2XX, 3XX). | 2403.00961v2 | 4 |
|
Figure 1:Comparison of Model Size vs. PubMedQA Accuracy in the
Reasoning-Required Setting:Our dense and sparse MediSwift models noticeably
outperform other fine-tuned language models≤\leq≤7B parameters, improving
the efficiency-accuracy pareto frontier. In particular, MediSwift-XL (1.21B)
achieves new state-of-the-art 76.8% accuracy at this size (i.e., being 5.8x
smaller than PMC-LlaMA). In addition, sparse pre-trained MediSwift-XL models
ats∈{50%,75%}𝑠percent50percent75s\in\{50\%,75\%\}italic_s ∈ { 50 % , 75 % }outperform other models at similar or larger size.
Additional details are provided in Table2and3. | 2403.00952v2 | 1 |
|
Figure 2:Comparison of pre-training loss curves for MediSwift
models: MediSwift-XL’s training loss reveals that at 50% sparsity, the
model’s performance closely mirrors that of its dense variant, with
negligible effects on training loss. At 75% sparsity, although the gap in
training loss widens, the sparse MediSwift-XL still outperforms the dense
MediSwift-Med, showcasing efficient learning even at higher sparsity
levels. | 2403.00952v2 | 2 |
|
Figure 3:Comparison of measured speedup versus theoretical speedup for
GPT-3 layer 12k×\times×12k matrix multiplication (MatMul) on the
Cerebras CS-2 system at various sparsity levels. This graph illustrates
the efficiency gains achieved through sparse computation, highlighting the
real-world performance relative to theoretical predictions. | 2403.00952v2 | 3 |
|
Figure 1:o.o.d.performance comparison between very large dropout, ensembles, and weight averaging methods on fourDomainBedtasks and two backbones (ResNet50andViT-L-16). Baseline results were obtained using plain fine-tuning with different hyperparameters.Weight averagingresults either average the model weights collected every 300 iterations along each fine-tuning trajectory or the final model weights of all fine-tuning trajectories as in(Ramé et al.,2022b).Ensembleresults average instead the model outputs. Finally,large dropoutresults were obtained like the baseline results but using a 90% dropout rate on the penultimate layer. Each box summarizes the results obtained with 1296 hyper-parameters combinations (no hyper-parameter selection). | 2403.00946v2 | 1 |
|
Figure 2:Effect of diverse dropout rates during fine-tuning. The besto.o.d.performances are attained using rates around or above 90%. A large dropout rate (e.g. 90%) reliably produces goodo.o.d.performance on all four tasks. | 2403.00946v2 | 2 |
|
Figure 3:Comparison of various dropout rates on the largerdomainnetdataset (586K examples), whose size approaches the pretraining dataset size (ImageNet, 1.2M examples). The optimal dropout rate falls to about 50%, a value comparable to the dropout rates traditionally used when training from scratch. | 2403.00946v2 | 3 |
|
Figure 3:Comparison of various dropout rates on the largerdomainnetdataset (586K examples), whose size approaches the pretraining dataset size (ImageNet, 1.2M examples). The optimal dropout rate falls to about 50%, a value comparable to the dropout rates traditionally used when training from scratch. | 2403.00946v2 | 4 |
|
Figure 4:Comparison of dropout rates when training aResNet50networkfrom scratchon thevlcsdataset. The optimal dropout rate falls to about zero. Dropout rates greater than 50% negatively impact both thei.i.d.and theo.o.d.performances. | 2403.00946v2 | 5 |
|
Figure 5:Incremental benefits achieved by constructing ensembles or by averaging the weights of models fine-tuned with very large dropouts. The baseline and dropout results are the same as those reported in Figure7. In contrast, the ensemble and weight averaging results are now obtained by averaging the output or the weights of models fine-tunedwith large dropouts. Ensemble and weight averaging techniques provide a marginalo.o.d.performance improvement onvlcsoroffice homeand a negligibleo.o.d.performance improvement onpacsorterra incognita. | 2403.00946v2 | 6 |
|
Figure 7:Performance comparison between very large dropout, ensembles, and weight averaging methods on fourDomainBedtasks andResNet50 #2backbone. The horizontal axis denotes thei.i.d.performance and the vertical axis theo.o.d.performance. Baseline results were obtained using plain fine-tuning with different hyperparameters (1296×\times×⚫).Weight averagingresults either average the model weights collected every 300 iterations along each fine-tuning trajectory (1296×\times×★★\bigstar★) or the final model weights of all fine-tuning trajectories (1×\times×\contourblack★★\bigstar★) as in(Ramé et al.,2022b).Ensembleresults average instead the model outputs (1296×\times×▲and 1×\times×\contourblack▲▲\blacktriangle▲). Finally,large dropoutresults were obtained like the baseline results but using a 90% dropout rate on the penultimate layer (1296×\times×◆). | 2403.00946v2 | 8 |
|
Figure 1:The key concept of distributedDNN. Left: LightweightDNNssuffer from performance loss due to the limited computational resources; Middle: Edge computing often results in intolerable latency due to communication overhead; Right: DistributedDNNsdeploy a smallheadmodel on the resource-limited device to achieve task-oriented compression and a largetailmodel on the server to decode compressed features and execute the rest of model. | 2403.00942v2 | 1 |
|
Figure 2:A general framework for entropy coding in neural data compression. It consists of an encoder-decoder structure to extract a compact latent representationz𝑧zitalic_zand a prior model to estimate the distributionPZ(z)subscript𝑃𝑍𝑧P_{Z}(z)italic_P start_POSTSUBSCRIPT italic_Z end_POSTSUBSCRIPT ( italic_z ). Left: During training, the prior model is jointly optimized with the encoder-decoder backbone. Right: During inference, the learned priorPZ(z)subscript𝑃𝑍𝑧P_{Z}(z)italic_P start_POSTSUBSCRIPT italic_Z end_POSTSUBSCRIPT ( italic_z )is used as side information for arithmetic encoding/decoding. | 2403.00942v2 | 2 |
|
Figure 5:Entropy models are sensitive to high-frequency features. From left to right: images from ImageNet validation set, bit rate maps and total variation maps. | 2403.00942v2 | 9 |
|
Figure 6:Entropy models are more vulnerable to perturbation in background region. From left to right: images from ImageNet validation set, bit rate comparison betweenPGD-E andPGD-Acc (PGD-E allocate less bit inbluearea and more bit inredarea). | 2403.00942v2 | 10 |
|
Figure 2.A taxonomy of the currently existing forms of bias (based on Figure 1 from(Richardson and Gilbert,2021)). | 2403.00935v1 | 2 |
|
Figure 1:A brief overview of our differentially private knowledge distillation framework, DistilDP, which can be broken down in three steps: (1) A pre-trained teacher model is fine-tuned on a private corpus using DP-SGD. (2) Next, the teacher model will generate synthetic data using control codes. (3) Finally, a pre-trained student model is fine-tuned on the DP synthetic data with the DP fine-tuned teacher for knowledge distillation. | 2403.00932v2 | 1 |
|
Figure 1:Overview of PRIME.(Left) Our learning framework leverages a set of pre-built behavior primitives to scaffold manipulation tasks. (Middle) Given task demonstrations, we use a trajectory parser to parse each demonstration into a sequence of primitive types (such as “push”, “grasp” and “place”) and their corresponding parametersxisubscript𝑥𝑖x_{i}italic_x start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT. (Right) With these parsed sequences of primitives, we use imitation learning to acquire a policy capable of predicting primitive types (such as “grasp”) and corresponding parametersx𝑥xitalic_xbased on observations. | 2403.00929v3 | 1 |
|
Figure 2:Method Overview.We develop a self-supervised data collection procedure that randomly executes sequences of behavior primitives in the environment. With the generated dataset, we train an IDM that maps an initial states𝑠sitalic_sand a final states′superscript𝑠′s^{\prime}italic_s start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPTfrom segments in task demonstrations to a primitive typep𝑝pitalic_pand corresponding parametersx𝑥xitalic_x. To derive the optimal primitive sequences, we build a trajectory parser capable of parsing task demonstrations into primitive sequences using the learned IDM. Finally, we train the policy using parsed primitive sequences. | 2403.00929v3 | 2 |
|
Figure 3:Simulated Tasks. We perform evaluations on three tasks from the RoboSuite simulator[zhu2020robosuite]. The first two,PickPlaceandNutAssembly, are from the RoboSuite benchmark, withNutAssemblyfeaturing less initial randomization than the original task. We introduce a third task,TidyUp, to study long-horizon tasks and test the inverse dynamics model’s generalization to unseen environments. We create four environment variants in this domain, denoted as (A, B, C, D).TidyUptask is designed in environment (D), and we collect human demonstrations forTidyUpin the same environment (D). To gauge the inverse dynamics model’s generalization capability, we train two IDMs: IDM-D, based solely on data from environment (D), and IDM-ABC, trained on data from environments (A, B, C). While IDM-D is our default model for experiments, we use IDM-ABC to evaluate generalization in unseen environments. | 2403.00929v3 | 3 |
|
Figure 4:Quantitative evaluation in three simulated tasks.Our method significantly outperforms state-of-the-art imitation learning approaches, with success rates surpassing 95% in all three tasks. | 2403.00929v3 | 4 |
|
Figure 5:Visualization of output primitive sequences from trajectory parser.For each task, we select five human demonstrations and visualize the segmented primitive sequences as interpreted by the trajectory parser. | 2403.00929v3 | 5 |
|
Figure 6:Real-world Tasks.InCleanUp-Bin, the robot must push the popcorn box to the serving area and place the butter inside the bin. InCleanUp-Stack, the robot must push the popcorn box to the serving area and stack the butter on top of the popcorn. | 2403.00929v3 | 6 |
|
Figure 1.An illustration of the search relevance problem. Using the semantic information of queries and products along with the behavioral relationships between them, the goal is to classify the degree of relevance for unseen query-product pairs (labeled with question marks) as exact, substitute, complement, or irrelevant. | 2403.00923v1 | 1 |
|
Figure 2.An overview of the PP-GLAM architecture. The model contains four modules: (i) Data Processing handles the pre-computation of graph neighborhoods and the de-noising of the product information, (ii) Model Training uses the training samples with query-product pairs and corresponding ground-truth labels to learn the parameters of the GBDT-based ensemble containing different language models and graph neural networks (GNNs), (iii) Model Selection module utilizes interpretable SHAP values to eliminate low-performing models based on the constraints of the inference setup, and (iv) Model Inference loads the selected set of models into memory for inference on new batches. | 2403.00923v1 | 2 |
|
Figure 3.Correlation between edges of the dataset. Note that dense signals such as impressions and clicks have a low correlation with target edges of exact, substitute, complement, and irrelevant, whereas sparser signals such as adds and purchases are highly correlated. Hence, we must carefully consider the reliability-availability trade-off in practice. | 2403.00923v1 | 3 |
|
Figure 1:Several pipelines of point cloud modeling.(a) denotes point-based methods with only local perception, including point-based methods, such as PointNet(Qi et al.2017a), PointNet++(Qi et al.2017b), PointMLP(Ma et al.2022), and PointNeXt(Qian et al.2022). (b) is the transformer-based method with global perception but quadratic computational cost, including Point Transformer(Zhao et al.2021)and Point-MAE(Pang et al.2022). (c) represents Mamba-based methods, which offer advantages of global modeling and linear computational complexity. | 2403.00762v4 | 1 |
|
Figure 2:The architecture of our proposed Point Cloud Mamba.PCM encoder consists of four stages, each comprising a geometric affine module and several mamba layers. Point downsampling is performed between stages. The decoder only consists of point interpolation, feature concatenation, and MLP. | 2403.00762v4 | 2 |
|
Figure 3:The consistent traverse serialization strategy.The 3-D point cloud data is voxelized and then serialized into a 1-D point sequence according to a predefined order. M represents the total number of points in the point cloud. With the permutation of x, y, and z coordinates, consistent traverse serialization has six variants. | 2403.00762v4 | 3 |
|
Figure 4:The order prompts.Different colors represent different serialization orders.Npsubscript𝑁𝑝N_{p}italic_N start_POSTSUBSCRIPT italic_p end_POSTSUBSCRIPTorder prompts are mapped to the same channel size as the features and then concatenated to the beginning and end of the input point sequence. | 2403.00762v4 | 4 |
|
Figure 1:Semantic-aware permutation. An assistant model segments the original training sentence into several semantic chunks. Then, we re-order the chunks (including original, permuting or reversal) with a certain probability. | 2403.00758v3 | 1 |
|
Figure 2.Relative European call option prices in the Heston model against the moneyness of the stock, compared to the COS and the DGM methods, for four different times to maturity withr=0.0𝑟0.0r=0.0italic_r = 0.0andη=0.1𝜂0.1\eta=0.1italic_η = 0.1,ρ=0.0𝜌0.0\rho=0.0italic_ρ = 0.0,κ=0.01𝜅0.01\kappa=0.01italic_κ = 0.01,V0=0.03subscript𝑉00.03V_{0}=0.03italic_V start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT = 0.03andλ=2.0𝜆2.0\lambda=2.0italic_λ = 2.0. | 2403.00746v1 | 4 |
|
Figure 5.Relative European call option prices in the lifted Heston model withn=20𝑛20n=20italic_n = 20variance processes against the moneyness of the stock, compared to the COS and the DGM methods, for four different times to maturity withr=0.0𝑟0.0r=0.0italic_r = 0.0,η=0.1𝜂0.1\eta=0.1italic_η = 0.1,ρ=0.0𝜌0.0\rho=0.0italic_ρ = 0.0,κ=0.01𝜅0.01\kappa=0.01italic_κ = 0.01,V0=0.01subscript𝑉00.01V_{0}=0.01italic_V start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT = 0.01andλ=2.0𝜆2.0\lambda=2.0italic_λ = 2.0. | 2403.00746v1 | 11 |
|
Figure 7.Relative European call option prices in the lifted Heston model withn=1𝑛1n=1italic_n = 1variance process against the moneyness of the stock, compared to the COS and the DGM methods, for four different times to maturity withr=0.0𝑟0.0r=0.0italic_r = 0.0,η=0.3𝜂0.3\eta=0.3italic_η = 0.3,ρ=−0.7𝜌0.7\rho=-0.7italic_ρ = - 0.7,κ=0.02𝜅0.02\kappa=0.02italic_κ = 0.02,V0=0.02subscript𝑉00.02V_{0}=0.02italic_V start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT = 0.02andλ=0.3𝜆0.3\lambda=0.3italic_λ = 0.3. | 2403.00746v1 | 15 |
|
Figure 9.Relative European call option prices in the lifted Heston model withn=5𝑛5n=5italic_n = 5variance processes against the moneyness of the stock, compared to the COS and the DGM methods, for four different times to maturity withr=0.0𝑟0.0r=0.0italic_r = 0.0,η=0.3𝜂0.3\eta=0.3italic_η = 0.3,ρ=−0.2𝜌0.2\rho=-0.2italic_ρ = - 0.2,κ=0.02𝜅0.02\kappa=0.02italic_κ = 0.02,V0=0.02subscript𝑉00.02V_{0}=0.02italic_V start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT = 0.02andλ=0.3𝜆0.3\lambda=0.3italic_λ = 0.3. | 2403.00746v1 | 19 |
|
Figure 3:A linear approximation to the attention probability is a particularly poor approximation in cases where one or both of the endpoints are in a saturated region of the softmax. Note that when varying only a single key, the softmax becomes a sigmoid of the dot product of that key and the query. | 2403.00745v1 | 8 |
|
Figure 4:Ranks ofc(n)𝑐𝑛c(n)italic_c ( italic_n )against ranks ofc^AtP(n)subscript^𝑐AtP𝑛\hat{c}_{\text{AtP}}(n)over^ start_ARG italic_c end_ARG start_POSTSUBSCRIPT AtP end_POSTSUBSCRIPT ( italic_n ), on Pythia-12B onCITY-PP. Both improvements to AtP reduce the number of false negatives (bottom right triangle area), where in this case most improvements come from the QK fix. Coloration indicates the maximum absolute difference in attention probability when comparingxcleansuperscript𝑥cleanx^{\text{clean}}italic_x start_POSTSUPERSCRIPT clean end_POSTSUPERSCRIPTand patching a given query or key. Many false negatives are keys and queries with significant maximum difference in attention probability, suggesting they are due to attention saturation as illustrated inFigure3. Output and value nodes are colored in grey as they do not contribute to the attention probability. | 2403.00745v1 | 9 |
|
Figure 5:True rank and rank of AtP estimates with and without GradDrop, using Pythia-12B on theCITY-PPdistribution withNeuronNodes. GradDrop provides a significant improvement to the largest neuron false negatives (red circles) relative to Default AtP (orange crosses). | 2403.00745v1 | 10 |
|
Figure 13:Cost of verified 100% recall curves, sweeping across models and settings forAttentionNodes | 2403.00745v1 | 58 |
|
Figure 14:True ranks against AtP*{}^{*}start_FLOATSUPERSCRIPT * end_FLOATSUPERSCRIPTranks on Pythia-12B using various metricsℒℒ\mathcal{L}caligraphic_L. The last row shows the effect in the denoising (rather than noising) setting; we speculate that the lower-right subplot (log-odds denoising) is similar to the lower-middle one (logit-diff denoising) because IOI produces a bimodal distribution over the correct and alternate next token. | 2403.00745v1 | 74 |
|
Figure 16:A comparison of edge-AtP variants across model sizes and prompt lengths. AtP* here is defined to include QKfix and MLPfix, but not GradDrops. The costs vary across several orders of magnitude for each setting.In the setting with full-MLP nodes, MLPfix carries substantial cost for short prompts, but barely matters for long prompts.In the neuron-nodes setting, MLPfix is costless. But GradDrops in that setting continues to impose a large cost; even though it doesn’t affect MLP→→\rightarrow→MLP edges, it does affect MLP→→\rightarrow→Q,K edges, which come out dominating the cost with QKfix. | 2403.00745v1 | 79 |
Subsets and Splits