image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
Figure 1:Standard speculative decoding versus our judge decoding strategy forLlama-3.1-8Bas draft andLlama-3.1-405Bas target. Accepted (rejected) tokens are highlighted in green (red). | 2501.19309v1 | 1 |
|
Figure 2:Left:Average number of generated tokens as a function of the number of draft tokensM𝑀Mitalic_MforLlama-8B/405Bwith standard and judge verification.Right:Number of accepted tokens on high-quality human text (top) and for both8B/405Band405B/8B(bottom), both standard SD. | 2501.19309v1 | 2 |
|
Figure 3:Left:Standard SD and our judge decoding whenGPT-4ois drafting andLlama-405Bis verifying. Green denotes accepted and red rejected tokens.Right:Number of accepted tokens forGPT-4oas draft andLlama-405Bas target for standard speculative and our judge verification. | 2501.19309v1 | 3 |
|
Figure 4:Two examples from our dataset. We highlight the incorrect tokens in the wrong answer in red. | 2501.19309v1 | 4 |
|
Figure 5:Left:ConditioningLlama-405Bon wrong outputs. The part of the assistant response in red was forced, while parts in green were generated freely.Right:Judge illustration wheresLsubscript𝑠𝐿s_{L}italic_s start_POSTSUBSCRIPT italic_L end_POSTSUBSCRIPTis the last token from the context𝒔𝒔\bm{s}bold_italic_sandc1,…,cMsubscript𝑐1…subscript𝑐𝑀c_{1},\dots,c_{M}italic_c start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT , … , italic_c start_POSTSUBSCRIPT italic_M end_POSTSUBSCRIPTare candidate tokens. Orange denotes embeddings, green denotes the LM-head output and red denotes the produced judgements. | 2501.19309v1 | 5 |
|
Figure 6:Benchmark results.Top:DraftLlama-8Band targetLlama-70B.Bottom:DraftLlama-8Band targetLlama-405B. We show top-K𝐾Kitalic_Kdecoding, standard SD forM=10𝑀10M=10italic_M = 10and our judge decoding forM=25𝑀25M=25italic_M = 25(striped). Notice that our judging method preserves accuracies very well, while top-K𝐾Kitalic_Kloses most performance. | 2501.19309v1 | 6 |
|
Figure 7:Illustration of mask creation in judge decoding. The decision mask resulting from the judge is combined with the standard mask from SD. Once both methods disagree, subsequent tokens get rejected automatically as usual, even if they were individually accepted. | 2501.19309v1 | 7 |
|
Figure 8:Illustration how combining the masks of judge decoding (JD) and standard speculative decoding (SD) results in the same reply but in less steps in the (rare) case that JD rejects a token that is actually the target token. | 2501.19309v1 | 8 |
|
Figure 9:More example prompts for SD forLlama-8BandLlama-405B.Left:Correct response getting rejected early under standard decoding, while judge decoding accepts a long continuation (but admittedly over-cautiously rejects later on).Right:Wrong response that gets rejected too early by standard decoding and correctly rejected later on by judge decoding (there are 5 roommates in total). | 2501.19309v1 | 9 |
|
Figure 10:More example prompts for speculative decoding forLlama-8BandLlama-405B.Left:Correct response getting rejected early under standard decoding, while judge decoding accepts a long continuation (but admittedly over-cautiously rejects later on).Right:Wrong response that gets rejected too early by standard decoding and correctly rejected later on by judge decoding (there are 5 roommates in total). | 2501.19309v1 | 10 |
|
Figure 11:Two example prompts from the subset ofwikipedia-summaries, along with the correspond verifications. | 2501.19309v1 | 11 |
|
Figure 12:Left:Llama-405Bcorrects the mistake by subtractingl[i]to double instead of triple.Right:Similar correction behaviour by pointing out that response so far is wrong. | 2501.19309v1 | 12 |
|
Figure 13:Left:Llama-405Bcorrectly points out that there is a bug as the index variable “j”” is not defined.Right:The model can also catch more subtle mistakes. Here the original list also gets modified as no copy was made, leading to wrong outputs. | 2501.19309v1 | 13 |
|
Figure 1:Illustration of the proposed Self-Enhanced Test-Time Scaling (SETS) method. SETS integrates Sampling, Self-Verify, and Self-Correct operations to efficiently scale test-time compute. | 2501.19306v2 | 1 |
|
Figure 2:Scaling law curves where the x-axis is the average number of output tokens and y-axis is the accuracy. Each point(x,y)𝑥𝑦(x,y)( italic_x , italic_y )in the curve corresponds to a hyperparameter settingθ∈Θ𝜃Θ\theta\in\Thetaitalic_θ ∈ roman_Θ.y𝑦yitalic_yis the optimal performance at the cost budgetx=H(θ)𝑥𝐻𝜃x=H(\theta)italic_x = italic_H ( italic_θ )(see Section4for details). We subsample the points (up to 6 within every x-tick interval) to make the markers less crowded. | 2501.19306v2 | 2 |
|
Figure 3:Scaling law curves for Trip Planning Hard and Easy tasks obtained with SETS vs. competing methods. The results for Meeting Planning are shown in AppendixD.6. | 2501.19306v2 | 3 |
|
Figure 4:The effect of allocating more compute to self-verification and self-correction for SETS (controlled by Max Number of Rounds) given a fixed computational budget (measured by Average Number of Output Tokens). The results for other datasets are shown in AppendixD.4. | 2501.19306v2 | 4 |
|
Figure 5:The effect of using different LLMs. The results for other datasets are shown in AppendixD.5. | 2501.19306v2 | 5 |
|
Figure 6:Scaling law curves where the x-axis is the average number of API calls and y-axis is the accuracy. Each point(x,y)𝑥𝑦(x,y)( italic_x , italic_y )in the curve corresponds to a hyperparameter settingθ∈Θ𝜃Θ\theta\in\Thetaitalic_θ ∈ roman_Θ.y𝑦yitalic_yis the optimal performance at the cost budgetx=H(θ)𝑥𝐻𝜃x=H(\theta)italic_x = italic_H ( italic_θ )(refer to Section4for the details). We subsample the points (up to 6 within every x-tick interval) to make the markers less crowded. | 2501.19306v2 | 6 |
|
Figure 7:Evaluate the self-verification performance of different models as we increase the number of self-verification samples. | 2501.19306v2 | 7 |
|
Figure 8:Evaluate the self-correction performance of different models as we increase the maximum number of self-correction rounds. The accuracy for00maximum number of rounds is the initial accuracy without self-correction. | 2501.19306v2 | 8 |
|
Figure 9:The effect of allocating more compute to self-verification and self-correction for SETS (controlled by Max Number of Rounds) given a fixed computational budget (measured by Average Number of Output Tokens). | 2501.19306v2 | 9 |
|
Figure 11:Estimated scaling law curves for Meeting Planning Hard and Easy tasks obtained with SETS vs. alternative methods. | 2501.19306v2 | 11 |
|
Figure 2:Reconstruction Loss of Model Trained with Full/Compressed Data on The Test Dataset (TOP 50). | 2501.19298v1 | 2 |
|
Figure 3:Trackingλ(t)superscript𝜆𝑡\lambda^{(t)}italic_λ start_POSTSUPERSCRIPT ( italic_t ) end_POSTSUPERSCRIPTfromDPS-MOZO+On, the average value across all test queries of the smallestλisubscript𝜆𝑖\lambda_{i}italic_λ start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT(y-axis) at thet𝑡titalic_t-th generation (x-axis). | 2501.19287v1 | 5 |
|
Figure 1:Evolution of the running time and objective on the MNIST dataset. Left: evolution as a function ofn𝑛nitalic_nfork=10𝑘10k=10italic_k = 10. Right: evolution as a function ofk𝑘kitalic_kforn=10000𝑛10000n=10000italic_n = 10000. The results for five competitors are reported: k-means++ (KM), FasterPAM (FP), FasterCLARA-5 (FC), BanditPAM++-2 (BP), OneBatchPAM (OBP) | 2501.19285v1 | 1 |
|
Figure 1:Proposed modelDepressionX. It predicts depression severity of a post and generates explanation with word- and sentence-level attentions and a knowledge subgraph. The prediction and the explanatory components are highlighted in yellow. | 2501.14985v1 | 1 |
|
Figure 3:Two-dimensional plots of post representations (generated by UMAP) before and after knowledge infusion | 2501.14985v1 | 3 |
|
(a)Left: learning rate, Right: number of heads in multihead attention (H) and number of heads in GAT (GH) | 2501.14985v1 | 7 |
|
(a)Left: learning rate, Right: number of heads in multihead attention (H) and number of heads in GAT (GH) | 2501.14985v1 | 8 |
|
Figure 1:Simulation performance of S4D-FT relative to the LSTM model across study watersheds in CONUS. Panel a: Spatial distribution of NSE skill scores, with red dots (positive NSE skill score) indicating S4D-FT outperformance and blue dots (negative NSE skill score) indicating LSTM outperformance. Panel b: Spatial distribution of KGE skill scores, following the same color scheme as Panel a. Panel c: Boxplots of NSE and KGE skill scores by state, ordered east to west, with red boxes for positive median scores and blue for negative medians. Rhode Island and Minnesota are excluded due to no study watersheds. | 2501.14980v1 | 1 |
|
Figure 1:GNNs with personalized prediction layer can empower existing GNNs for moreaccurate,personalized,robust, anddeepernode classification (e.g., classifying users as conservative (purple) or liberal (green)). For example, while some nodes benefit from their neighbors, others may be better off by predicting using only their attributes. | 2501.14964v1 | 1 |
|
Figure 2:GNN with node-optimal prediction layer trained using MetSelect. Labeled nodes are denoted in purple and green while gray denotes unlabeled nodes. | 2501.14964v1 | 2 |
|
Figure 4:Comparison of fine-tuned LLaMA-2 models on generating improper responses, illustrating the impact of the custom loss function, augmented prompts, and training epochs on model performance. | 2501.14960v1 | 4 |
|
Figure 6:Comparison of the proposed Fine-tuned LLama-3.1 7B (20 epochs) against the baseline models, LLama-2 7B, Llama-3.1 8B, Falcon 7B and Mistral 7B evaluated on the combined network dataset. | 2501.14960v1 | 6 |
|
Figure 7:Model performance in inferring (a) system loss and (b) system voltages(mean absolute Error (MAE) is almost 0). | 2501.14960v1 | 7 |
|
Figure 8:Comparison of the proposed Fine-tuned Llama-3.1 7B trained over 10, 20 and 30 epochs over the number of improper outputs. | 2501.14960v1 | 8 |
|
Figure 9:Comparison of the proposed Fine-tuned Llama-3.1 7B trained over 10, 20 and 30 epochs over the number of suboptimal config. ouputed. | 2501.14960v1 | 9 |
|
Figure 3.The role of multiplicity in brute-force search for bi-level and constrained optimization problems in machine learning. | 2501.14959v1 | 3 |
|
(b)After applying rule(x+y)−z→x+(y−z)→𝑥𝑦𝑧𝑥𝑦𝑧(x+y)-z\rightarrow x+(y-z)( italic_x + italic_y ) - italic_z → italic_x + ( italic_y - italic_z ) | 2501.14951v1 | 3 |
|
Figure 2:Overview of E-Gen and the cluster-based training framework. Equivalent expressions of each initial expression are generated using an e-graph-based approach, forming clusters of equivalent expressions. The seq2seq model is trained on equivalent expression pairs, while the contrastive learning model is trained on triplets, with each triplet containing a reference expression𝒙𝒙\boldsymbol{x}bold_italic_x, an equivalent positive sample𝒙+superscript𝒙\boldsymbol{x}^{+}bold_italic_x start_POSTSUPERSCRIPT + end_POSTSUPERSCRIPT, and a non-equivalent negative sample𝒙−superscript𝒙\boldsymbol{x}^{-}bold_italic_x start_POSTSUPERSCRIPT - end_POSTSUPERSCRIPT.𝒛𝒛\boldsymbol{z}bold_italic_zis the latent space representation of the corresponding input𝒙𝒙\boldsymbol{x}bold_italic_x. | 2501.14951v1 | 7 |
|
Figure 3:Visualization of representation vectors for 17 different single-operator mathematical expressions and their equivalent forms using our method (left) andSemEmb(right). t-SNE(Van der Maaten and Hinton,2008)is applied to reduce the dimensionality of the embeddings from 512 to 2. | 2501.14951v1 | 8 |
|
Figure 4:Example of the formula retrieval with seq2seq model. The query expression iscos(x)𝑥\cos(x)roman_cos ( italic_x )in darker blue, and the rest are candidates. Top-4444expressions are retrieved. t-SNE is applied to reduce the dimensionality of the embeddings from 512 to 2. | 2501.14951v1 | 9 |
|
Figure 5:Two mistake detection conversation examples in GPT-4o prompt. The mistakes in the derivation are in red. In the left one, GPT-4o successfully detects the mistake step, but also reports a false positive mistake step. In the right one, GPT-4o fails to detect the mistake in the derivation at all. | 2501.14951v1 | 10 |
|
Figure 1:MATCHA formatching anything.We visualize geometric, semantic and temporal correspondences established by MATCHA, using a single feature descriptor. | 2501.14945v1 | 1 |
|
Figure 2:Heatmap of features from DINOv2, DIFT, and MATCHA.Given a querypointfrom the source image (1st column), DINOv2 features give more accurate correspondences on single object (1st and 2nd row) but struggle when multiple instances of the same class (3rd row) or similar structures (4th row) exist. Both geometric and semantic features of DIFT perform reversely. By unifying knowledge in the three foundation features, MATCHA produces more accurate and reliable correspondences. | 2501.14945v1 | 2 |
|
Figure 3:Architecture of MATCHA.Given an RGB image, MATCHA produces a single feature for geometric, semantic and temporal matching with nearest neighbor searching. MATCHA is built on top of stable diffusion (SD) models[53]and DINOv2[44]. Specifically, original geometric and semantic features extracted from SD are first fused dynamically with a transformer[64]consists of self and cross attention blocks. In this dynamic fusion process, both geometric and semantic features are augmented with each other which are supervised with corresponding ground-truth signals in the training process. Then, augmented geometric and semantics features along with DINOv2 feature are unified statically via concatenations into a single feature formatching anything. | 2501.14945v1 | 3 |
|
Figure 4:Geometric Matching on HPatches. We report Mean Matching Accuracy (MMA) at error thresholds ranging from 1-10 pixel. Concrete and dash lines denote methods with and without supervision, respectively. | 2501.14945v1 | 4 |
|
Figure 5:Visualization of temporal matches on TapVID-Davis[12].Here we visualize several challenging cases for exstablishing temporal correspondences, where MATCHA generally achieves the best performance in handling extreme scale and viewpoint changes, as well as scenes with multiple similar instances. (DIFT* is the adapted DIFT where we use its concatenated semantic and geometric feature for temporal matching for better performance. ) | 2501.14945v1 | 5 |
|
Figure 6:Visualization of heatmap.Given a source point (top), we visualize the heatmap and predicted matches of MASt3R.E[30], DISK[63], DINOv2[44], DIFT[60], DIFT.S (fully supervised version of DIFT), and our models MATCHA-Light and MATCHA. | 2501.14945v1 | 6 |
|
Figure 7:Geometric matches on outdoor scenes.We visualize the inliers after RANSAC of MASt3R.E[30], DISK[63], DINOv2[44], DIFT[60], DIFT.S (fully supervised version of DIFT), and our models MATCHA-Light and MATCHA. DISK produces many inliers on local patches but is not robust to repetitive structures. MASt3R and DINOv2 focus more on structures and give close performance. DIFT works better than DINOv2 especially on regions with rich textures. With geometric supervision, DIFT.S improves the performance of DIFT. MATCHA-Light is able to find correct matches from both local patches and structures because of dynamic fusion and this ability is further enhanced by fusing features of DINOv2. | 2501.14945v1 | 7 |
|
Figure 8:Geometric matches on indoor scenes.We visualize the inliers after RANSAC of MASt3R.E[30], DISK[63], DINOv2[44], DIFT[60], DIFT.S (fully supervised version of DIFT), and our models MATCHA-Light and MATCHA. Almost all previous methods fail to find sufficient inliers on scenes with repetitive structures except MATCHA which fuses both low and high-level information. Additionally, MATCHA is able to produce more inliers in scenes with rich textures (middle column). | 2501.14945v1 | 8 |
|
Figure 9:Semantic matches on bus category.We visualize theinliersandoutliersof MASt3R.E[30], DISK[63], DINOv2[44], DIFT[60], DIFT.S (fully supervised version of DIFT), and our models MATCHA-Light and MATCHA. | 2501.14945v1 | 9 |
|
Figure 10:Semantic matches on plant category.We visualize theinliersandoutliersof MASt3R.E[30], DISK[63], DINOv2[44], DIFT[60], DIFT.S (fully supervised version of DIFT), and our models MATCHA-Light and MATCHA. | 2501.14945v1 | 10 |
|
Figure 11:Semantic matches on sheep category.We visualize theinliersandoutliersof MASt3R.E[30], DISK[63], DINOv2[44], DIFT[60], DIFT.S (fully supervised version of DIFT), and our models MATCHA-Light and MATCHA. | 2501.14945v1 | 11 |
|
Figure 12:Semantic matches on chair category.We visualize theinliersandoutliersof MASt3R.E[30], DISK[63], DINOv2[44], DIFT[60], DIFT.S (fully supervised version of DIFT), and our models MATCHA-Light and MATCHA. | 2501.14945v1 | 12 |
|
Figure 13:Semantic matches on motorbike category.We visualize theinliersandoutliersof MASt3R.E[30], DISK[63], DINOv2[44], DIFT[60], DIFT.S (fully supervised version of DIFT), and our models MATCHA-Light and MATCHA. | 2501.14945v1 | 13 |
|
Figure 14:Temporal matches on goldfish sequence.We visualize theinliersandoutliersof MASt3R.E[30], DISK[63], DINOv2[44], DIFT.Uni[60], DIFT.S.Uni (fully supervised version of DIFT), and our models MATCHA-Light and MATCHA. | 2501.14945v1 | 14 |
|
Figure 15:Temporal matches on horsejumphigh sequence.We visualize theinliersandoutliersof MASt3R.E[30], DISK[63], DINOv2[44], DIFT.Uni[60], DIFT.S.Uni (fully supervised version of DIFT), and our models MATCHA-Light and MATCHA. | 2501.14945v1 | 15 |
|
Figure 16:Temporal matches on soapbox sequence.We visualize theinliersandoutliersof MASt3R.E[30], DISK[63], DINOv2[44], DIFT.Uni[60], DIFT.S.Uni (fully supervised version of DIFT), and our models MATCHA-Light and MATCHA. | 2501.14945v1 | 16 |
|
Figure 17:Temporal matches on scooterblack sequence.We visualize theinliersandoutliersof MASt3R.E[30], DISK[63], DINOv2[44], DIFT.Uni[60], DIFT.S.Uni (fully supervised version of DIFT), and our models MATCHA-Light and MATCHA. | 2501.14945v1 | 17 |
|
Figure 1:Visualization of an example where context influences human judgments on whether it is safe to respond to a user’s query. Context is formalized using CI parameters: sender, recipient, and transmission principle. Left: No context; Middle: Safe context; Right: Unsafe context. More context examples can be found in the tutorial in AppendixD.1. | 2501.14940v2 | 1 |
|
Figure 2:Data creation pipeline for CASE-Bench which sequentially executes query selection, automatic context generation and manual revision stages. Annotators are involved in each stage. | 2501.14940v2 | 2 |
|
Figure 3:Visualization of Kruskal-Wallis test results across 45 categories in the CASE-Bench dataset. The chart distinguishes between significant and non-significant categories, with bars representing the average K-W statistic values. Categories labelled as ”Non-significant” are displayed with a hatched pattern. Category “Child-related Crimes” is excluded as detailed in §4.1. | 2501.14940v2 | 3 |
|
Figure 4:Plot of correlation between LLM safety judgements and safety ratings given by the crowd of annotators. As multiple dots may overlay on each other, the density of the dots is also indicated by the colour map in the background where darker means denser. | 2501.14940v2 | 4 |
|
Figure 5:Recall rates for safe and unsafe contexts with different subsets of CI parameters for Llama-3 (upper) and GPT-4o-mini (middle) and Claude-3.5 (bottom). S denotes sender, R denotes recipient, TP denotes the transmission principle. | 2501.14940v2 | 5 |
|
Figure 6:The custom user interface we developed for Amazon MTurk annotators, enabling them to provide judgments conveniently and efficiently. | 2501.14940v2 | 6 |
|
Figure 7:The required number of annotators per task determined through power analysis using the G*Power. The specific values used in the analysis are justified in the §3.2. | 2501.14940v2 | 7 |
|
Figure 8:Overview of the two statistical methods used in the paper: ① the Kruskal-Wallis test, performed on atask-by-taskbasis, and ② the z-test, applied across the entire set of tasks. | 2501.14940v2 | 8 |
|
Figure 9:Visualization of Human Annotations for Tasks in the “Religion Promotion” Category. Tasks with queries outlined in blue indicate a significant K-W test result. The figure is based on the responses of 21 annotators, showing the proportion of those who selected either “Should respond” or “Should refuse”. | 2501.14940v2 | 9 |
|
Figure 10:Visualization of Human Annotations for Tasks in the “Social Stereotype Promotion” Category. Tasks with queries outlined in blue indicate a significant K-W test result. The figure is based on the responses of 21 annotators, showing the proportion of those who selected either “Should respond” or “Should refuse”. | 2501.14940v2 | 10 |
|
Figure 1:This figure visualizes the sample adjacency matrix for each model atn=5000𝑛5000n=5000italic_n = 5000andK=20𝐾20K=20italic_K = 20, the sample community scores fork∈[1,20]𝑘120k\in[1,20]italic_k ∈ [ 1 , 20 ], and the principal graph encoder embedding. | 2501.14939v1 | 1 |
|
Figure 2:This figure displays the average sample community score, the average principal community detection accuracy, and the average vertex classification error using the embedding, based on100100100100Monte-Carlo replicates with increasingn𝑛nitalic_n. P-GEE denotes the principal graph encoder embedding, and GEE denotes the original graph encoder embedding. | 2501.14939v1 | 2 |
|
Figure 1:Performance of all baselines in four simulation scenarios described in Section4.2. The red vertical lines correspond to target coverage, and the blue vertical lines correspond to the optimal interval width. Here CD stands for Algorithm1, WCP stands for weighted conformal prediction, and CM stands for conformal meta-learners. | 2501.14933v1 | 1 |
|
Figure 1:This figure shows the graph covariance between the baseline feature and the other simulated binary features across all timestamps. | 2501.14932v1 | 1 |
|
Figure 2:This figure visualizes the graph covariance matrix for the years 2003, 2009, 2015, and 2021 from the CTDC data. | 2501.14932v1 | 2 |
|
Figure 3:This figure visualizes the graph covariance and its changes over time for several noteworthy pairs of features. | 2501.14932v1 | 3 |
|
Figure E1:Same experiments as Figure1but using distance correlation (top two panels) and Hilbert-Schmidt independence criterion (bottom two panels). | 2501.14932v1 | 4 |
|
Figure 1:Distribution of hypervolume values by approach and baseline across the three data splits. Higher values indicate better performance. | 2502.00034v1 | 1 |
|
(a)Distribution ofmaxρn−1subscript𝜌𝑛1\max\rho_{n-1}roman_max italic_ρ start_POSTSUBSCRIPT italic_n - 1 end_POSTSUBSCRIPTby approach and baseline across the three data splits. | 2502.00034v1 | 2 |
|
(a)Distribution ofmaxρn−1subscript𝜌𝑛1\max\rho_{n-1}roman_max italic_ρ start_POSTSUBSCRIPT italic_n - 1 end_POSTSUBSCRIPTby approach and baseline across the three data splits. | 2502.00034v1 | 3 |
|
(b)Number of solved days by approach and baseline for each data split. A day is considered solved if at least one of the produced plans bringsmaxρn−1subscript𝜌𝑛1\max\rho_{n-1}roman_max italic_ρ start_POSTSUBSCRIPT italic_n - 1 end_POSTSUBSCRIPTunder the threshold of 1. | 2502.00034v1 | 4 |
|
Figure 3:Depiction of part of the grid used in our experiments. The entire grid contains 1659 substations, 1338 power lines and 1716 injections. | 2502.00034v1 | 5 |
|
Figure 4:Corresponding "N switching" objective average for best plans with bestmaxρn−1subscript𝜌𝑛1\max\rho_{n-1}roman_max italic_ρ start_POSTSUBSCRIPT italic_n - 1 end_POSTSUBSCRIPTdepicted in Fig.2. | 2502.00034v1 | 6 |
|
Figure 1:Illustration of the proposed temporal-aware segmentation model.(A)The baseline UNet (2D or 3D) processes inputs frame by frame, ignoring temporal information (time-independent).(B)A conventional temporal UNet incorporates time by using Conv3D for 2D+t[9]or Conv4D for 3D+t[10]data.(C)The proposed model improves temporal integration by combining Conv2D (for 2D+t) or Conv3D (for 3D+t) with a novelTemporal Attention Module (TAM). The TAM employs multi-headed cross-time attention to refine motion features across frames, boosting segmentation quality while maintaining computational efficiency comparable to the model in(B)(Table1). | 2501.14929v1 | 1 |
|
Figure 2:The proposed model enhances temporal integration by coupling our novelTAMwith either a Conv2D (for 2D+t) or Conv3D (for 3D+t) network. In (A), TAM is integrated into the transformer-based UNetR model, while (B) demonstrates its integration with the CNN-based UNet. The TAM (as shown in (C)) leverages multi-headed cross-time attention to refine motion features across frames, improving segmentation quality without the computational overhead typically incurred by adding extra dimensions to the input tensors[9,10](as shown in Fig.1and Table1). | 2501.14929v1 | 2 |
|
Figure 4:Distribution plots of HD and PIA metrics of different segmentation configurations in Table2. | 2501.14929v1 | 6 |
|
Figure 5:The improvement in segmentation quality is attributed to the incorporation of motion, with the cost-efficient C3 and C4 configurations outperforming the computationally intensive C2 setup. | 2501.14929v1 | 7 |
|
Figure 7:Performance comparison of our motion-aware segmentation model with other models using the same FCN8s backbone architecture across varying image quality levels. | 2501.14929v1 | 11 |
|
Figure 9:Assessing the temporal consistency of segmentation results using FCN8s, both without motion and with motion integrated through our novel TAM, on the publicly available CAMUS dataset. | 2501.14929v1 | 16 |
|
Figure 1:Decomposing a target network’s parameters into parameter components that are faithful, minimal, and simple. | 2501.14926v3 | 1 |
|
Figure 2:Top:Step 1: Calculating parameter component attributionsAc(x)subscript𝐴𝑐𝑥A_{c}(x)italic_A start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT ( italic_x ).Bottom:Step 2: Optimizing minimality lossℒminimalitysubscriptℒminimality\mathcal{L}_{\text{minimality}}caligraphic_L start_POSTSUBSCRIPT minimality end_POSTSUBSCRIPT. | 2501.14926v3 | 2 |
|
Figure 5:The architecture of our Toy Model of Compressed Computation using a1111-layer residual MLP. We fixWEsubscript𝑊𝐸W_{E}italic_W start_POSTSUBSCRIPT italic_E end_POSTSUBSCRIPTto be a randomly generated matrix with unit norm rows, andWU=WE⊤subscript𝑊𝑈superscriptsubscript𝑊𝐸topW_{U}={W_{E}}^{\top}italic_W start_POSTSUBSCRIPT italic_U end_POSTSUBSCRIPT = italic_W start_POSTSUBSCRIPT italic_E end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⊤ end_POSTSUPERSCRIPT. | 2501.14926v3 | 5 |
|
Figure 7:Output of multiple APD forward passes with one-hot inputx42=1subscript𝑥421x_{42}=1italic_x start_POSTSUBSCRIPT 42 end_POSTSUBSCRIPT = 1over10101010k samples, where half of the parameter components are ablated in each run. Purple lines show ‘scrubbed’ runs (parameter component corresponding to input index42424242is preserved), while green lines show ‘anti-scrubbed’ runs (component42424242is among those ablated). The target model output is shown in blue, which is almost identical to the output on the APD sparse forward pass (i.e. APD (top-k𝑘kitalic_k)). In this plot we only show the MLP output for clearer visualization. The embedding matrices are not decomposed and thus the residual stream contribution does not depend on APD components. | 2501.14926v3 | 7 |
|
Figure 8:MSE losses of the APD model on the sparse forward pass (“top-k𝑘kitalic_k") and the APD model when ablating half (50505050) of its parameter components (“scrubbed" when none of the components responsible for the active inputs are ablated and “anti-scrubbed" when they are ablated). The gray line indicates the loss for a model that uses one monosemantic neuron per input feature. The dashed colored lines are the mean MSE losses for each type of run. | 2501.14926v3 | 8 |
|
Figure 9:The architecture of our Toy model of Cross-Layer Distributed representations using a 2-layer residual MLP. We fixWEsubscript𝑊𝐸W_{E}italic_W start_POSTSUBSCRIPT italic_E end_POSTSUBSCRIPTto be a randomly generated matrix with unit norm rows, andWU=WETsubscript𝑊𝑈superscriptsubscript𝑊𝐸𝑇W_{U}={W_{E}}^{T}italic_W start_POSTSUBSCRIPT italic_U end_POSTSUBSCRIPT = italic_W start_POSTSUBSCRIPT italic_E end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_T end_POSTSUPERSCRIPT. | 2501.14926v3 | 9 |
|
Figure 11:OptimizingℒminimalityLpsuperscriptsubscriptℒminimalitysubscript𝐿𝑝\mathcal{L}_{\text{minimality}}^{L_{p}}caligraphic_L start_POSTSUBSCRIPT minimality end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_L start_POSTSUBSCRIPT italic_p end_POSTSUBSCRIPT end_POSTSUPERSCRIPT | 2501.14926v3 | 11 |
|
Figure 12:Hand-coded gated function model: The four functionsfi(x)subscript𝑓𝑖𝑥f_{i}(x)italic_f start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT ( italic_x )implemented by the
hand-coded gated function model (solid lines), and the outputs
of the top-k𝑘kitalic_kforward pass of the APD-decomposed
model (dashed lines). The APD model almost perfectly
matches the hand-coded network. | 2501.14926v3 | 12 |
|
Figure 13:The parameters of the hand-coded gated function model decomposed into parameter components.Leftmost panel:The hand-coded network parameters, colored by the unique functionsFi(x)subscript𝐹𝑖𝑥F_{i}(x)italic_F start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT ( italic_x ).Other panels:The parameter components identified by APD, coloured by the function they correspond to in the target model, or purple if the weight is zero in the target model. | 2501.14926v3 | 13 |
|
Figure 14:Output of the1111-layer residual MLP target model compared to true labels for a single active input.Left: Output at all output indices for single one-hot inputx42=1subscript𝑥421x_{42}=1italic_x start_POSTSUBSCRIPT 42 end_POSTSUBSCRIPT = 1.Right: Output at indexj=42𝑗42j=42italic_j = 42for inputs withx42∈[0,1]subscript𝑥4201x_{42}\in[0,1]italic_x start_POSTSUBSCRIPT 42 end_POSTSUBSCRIPT ∈ [ 0 , 1 ]andxj=0subscript𝑥𝑗0x_{j}=0italic_x start_POSTSUBSCRIPT italic_j end_POSTSUBSCRIPT = 0forj≠42𝑗42j\neq 42italic_j ≠ 42. | 2501.14926v3 | 14 |
Subsets and Splits