image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
Figure 8:Visualization results of different CNN and Transformer hybrid networks (CvT, CMT, and SCB) and a network combining different convolutional kernel sizes (RepLKNet) on the SEN12 dataset. | 2501.10197v1 | 8 |
|
Figure 1:Organization of the contributions to this thesis. The paper highlighted in gray is not part of this habilitation thesis. However, it may be advisable to read it, as it discusses in detail the foundations of many concepts that are important for this thesis, in particular decision making based on preference systems. Apart from this, the graph is to be understood as follows: If a one-sided arrow leads from Contribution A to Contriution B, then B builds (in part) on A. If, on the other hand, A and B are connected by a double arrow, they relate in terms of content, without mathematically (in parts) building directly on each other. | 2501.10195v1 | 1 |
|
Figure 3:The figure shows an exemplary credal set on a finite set of statesS={s1,s2,s3}𝑆subscript𝑠1subscript𝑠2subscript𝑠3S=\{s_{1},s_{2},s_{3}\}italic_S = { italic_s start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT , italic_s start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT , italic_s start_POSTSUBSCRIPT 3 end_POSTSUBSCRIPT }. You can see directly that this can be identified (for finiteS𝑆Sitalic_S) with a convex polyhedron. We repeatedly take advantage of this fact in some of the publications to solve various problems using (mixed-integer) linear programming. | 2501.10195v1 | 3 |
|
Figure 5:The diagram shows a schematic comparison of the two different types of regularization in preference systems. Note that the identical diagram can be found inJansen et al. (2023b, Section 3). | 2501.10195v1 | 5 |
|
Figure 2.An example Lite-GPU deployment. Each NVIDIA H100 GPU is replaced with four Lite-GPUs, featuring better hardware yield and higher bandwidth-to-compute. | 2501.10187v1 | 2 |
|
(a)Prompt prefill. All configurations perform similarly.
As the model sizes grow, the “Lite” cluster underperforms due to increased collectives causing network bottleneck.
Increasing the network bandwidth compensates the increased network demand, overclocking improves performance further as prefill workloads are compute-bound. | 2501.10187v1 | 3 |
|
(a)Prompt prefill. All configurations perform similarly.
As the model sizes grow, the “Lite” cluster underperforms due to increased collectives causing network bottleneck.
Increasing the network bandwidth compensates the increased network demand, overclocking improves performance further as prefill workloads are compute-bound. | 2501.10187v1 | 4 |
|
(b)Decode. As model sizes and thus the number of required GPUs grow, the “Lite” cluster underperforms due to increased memory access intensities.
The degredation is worse with GPT-3 due to it having more KV-heads resulting in propotionally longer memory-bound stages.
As Lite-GPUs utilize their shoreline for more memory bandwidth, performance improves and exceeds the current H100 cluster. | 2501.10187v1 | 5 |
|
Figure 2:Distribution of label contribution to P@5 in Eurlex-4K and Wiki10-31K before and after applying weighting. | 2501.10179v1 | 2 |
|
Figure 2:Language Reduction (on the left) gives the XLM-R2 model that goes through Domain Adaptation (on the right) and brings the XLM-R2-ID model. Diagram fromPavlova and Makhlouf (2024)with permission. | 2501.10175v1 | 2 |
|
Figure 3:Comparison of the performance of the retrieval models trained from XLM-RBaseand XLM-R2-ID for MRR@10 in English. | 2501.10175v1 | 3 |
|
Figure 5:Comparison of the performance of the retrieval models trained from XLM-RBaseand XLM-R2-ID for Recall@100 in English. | 2501.10175v1 | 4 |
|
Figure 1:Loss function (10) forℛ+superscriptℛ\mathcal{R}^{+}caligraphic_R start_POSTSUPERSCRIPT + end_POSTSUPERSCRIPTdepending onλ^^𝜆\hat{\lambda}over^ start_ARG italic_λ end_ARG. Gray markers represent the numerical values of the loss function, determined with Alg.1. The green dashed line shows the upper bound (16) and the blue dashed line shows the lower bound (22). | 2501.10173v1 | 1 |
|
Figure 2:Loss function (10) forℛ×superscriptℛ\mathcal{R}^{\times}caligraphic_R start_POSTSUPERSCRIPT × end_POSTSUPERSCRIPTdepending onλ^^𝜆\hat{\lambda}over^ start_ARG italic_λ end_ARG. Gray markers represent the numerical values of the loss function, determined with Alg.1. The green dashed line shows the upper bound (26) and the blue dashed line shows the lower bound (33). | 2501.10173v1 | 5 |
|
Figure 3:Loss function (10) forℛ∗superscriptℛ\mathcal{R}^{*}caligraphic_R start_POSTSUPERSCRIPT ∗ end_POSTSUPERSCRIPTdepending onλ^^𝜆\hat{\lambda}over^ start_ARG italic_λ end_ARG. Gray markers represent the numerical values of the loss function, determined with Alg.1. The green dashed line shows the upper bound (44) and the blue dashed line shows the lower bound (54). | 2501.10173v1 | 9 |
|
Figure 4:Loss function (10) depending onλ^^𝜆\hat{\lambda}over^ start_ARG italic_λ end_ARG. Markers represent the numerical values of the loss function, determined with Alg.1, black for theℛ∗superscriptℛ\mathcal{R}^{*}caligraphic_R start_POSTSUPERSCRIPT ∗ end_POSTSUPERSCRIPT-RS and gray for theR×superscript𝑅R^{\times}italic_R start_POSTSUPERSCRIPT × end_POSTSUPERSCRIPT-RS. The green line in the middle and right plot shows the upper bound (44). The gray dashed line shows the lower bound (54) of theℛ∗superscriptℛ\mathcal{R}^{*}caligraphic_R start_POSTSUPERSCRIPT ∗ end_POSTSUPERSCRIPT-RS and the blue dashed line shows the lower bound (33) of theℛ×superscriptℛ\mathcal{R}^{\times}caligraphic_R start_POSTSUPERSCRIPT × end_POSTSUPERSCRIPT-RS. | 2501.10173v1 | 13 |
|
Figure 5:Left and middle plot: loss function (10) forℛ#superscriptℛ#\mathcal{R}^{\#}caligraphic_R start_POSTSUPERSCRIPT # end_POSTSUPERSCRIPTdepending onλ^^𝜆\hat{\lambda}over^ start_ARG italic_λ end_ARG. Gray markers represent the numerical values of the loss function, determined with Alg.1. The green dashed line shows the upper bound (58) and the blue dashed line shows the lower bound (67). Right plot: visualization for Eqs. (60) and (68). | 2501.10173v1 | 16 |
|
Figure 6:Left plot: lower bound (22) and upper bound (16) of theℛ+superscriptℛ\mathcal{R}^{+}caligraphic_R start_POSTSUPERSCRIPT + end_POSTSUPERSCRIPTloss function depending onν𝜈\nuitalic_νforλ0=2subscript𝜆02\lambda_{0}=2italic_λ start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT = 2andλ^=100^𝜆100\hat{\lambda}=100over^ start_ARG italic_λ end_ARG = 100. Right plot: markers show the relative loss function (77) forλ0=2subscript𝜆02\lambda_{0}=2italic_λ start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT = 2and solid lines represent the corresponding lower bound (79). | 2501.10173v1 | 20 |
|
Figure 7:Markers with dashed lines: relative loss function (80) of theℛ×superscriptℛ\mathcal{R}^{\times}caligraphic_R start_POSTSUPERSCRIPT × end_POSTSUPERSCRIPT-RS depending onλ^^𝜆\hat{\lambda}over^ start_ARG italic_λ end_ARG. Values were determined numerically with Alg.1. Solid lines: upper bound of the relative loss function (4.2). | 2501.10173v1 | 23 |
|
Figure 8:Markers with dashed lines: relative loss function (86) of theℛ∗superscriptℛ\mathcal{R}^{*}caligraphic_R start_POSTSUPERSCRIPT ∗ end_POSTSUPERSCRIPT-RS depending onλ^^𝜆\hat{\lambda}over^ start_ARG italic_λ end_ARG. Values were determined numerically with Alg.1. Solid lines: upper bound of the relative loss function. | 2501.10173v1 | 27 |
|
Figure 9:Markers with dashed lines: relative loss function (89) of theℛ#superscriptℛ#\mathcal{R}^{\#}caligraphic_R start_POSTSUPERSCRIPT # end_POSTSUPERSCRIPT-RS depending onλ^^𝜆\hat{\lambda}over^ start_ARG italic_λ end_ARG. Values were determined numerically with Alg.1. Solid lines: lower bound of the relative loss function (4.3). | 2501.10173v1 | 31 |
|
Figure 1.Sketch of the optimal transport problem, illustrating the mapping of a domain𝒳𝒳\mathcal{X}caligraphic_Xonto another domain𝒴𝒴\mathcal{Y}caligraphic_Y. | 2501.10162v1 | 1 |
|
Figure 2.Transport of a constant density from a unit disk domain to an ellipse domain.
Visualization of uniformly sampled source points and their image through exact and approximate transport maps. | 2501.10162v1 | 2 |
|
Figure 3.Transport of a constant density from a unit disk domain to an ellipse domain.
Visualization of the transport map∇uNN∇subscript𝑢𝑁𝑁\nabla u_{NN}∇ italic_u start_POSTSUBSCRIPT italic_N italic_N end_POSTSUBSCRIPT. | 2501.10162v1 | 3 |
|
Figure 4.Transport of a constant density from a unit disk domain to an ellipse domain.
Top left: firstx𝑥xitalic_x-coordinate of the transport map∇uNN∇subscript𝑢𝑁𝑁\nabla u_{NN}∇ italic_u start_POSTSUBSCRIPT italic_N italic_N end_POSTSUBSCRIPT;
Top right: Pointwise absolute error|∇uNN,x−∇uex,x|∇subscript𝑢𝑁𝑁𝑥∇subscript𝑢𝑒𝑥𝑥\left|\nabla u_{NN,x}-\nabla u_{ex,x}\right|| ∇ italic_u start_POSTSUBSCRIPT italic_N italic_N , italic_x end_POSTSUBSCRIPT - ∇ italic_u start_POSTSUBSCRIPT italic_e italic_x , italic_x end_POSTSUBSCRIPT |;
Bottom left: secondy𝑦yitalic_y-coordinate of the transport map∇uNN∇subscript𝑢𝑁𝑁\nabla u_{NN}∇ italic_u start_POSTSUBSCRIPT italic_N italic_N end_POSTSUBSCRIPT;
Top right: Pointwise absolute error|∇uNN,y−∇uex,y|∇subscript𝑢𝑁𝑁𝑦∇subscript𝑢𝑒𝑥𝑦\left|\nabla u_{NN,y}-\nabla u_{ex,y}\right|| ∇ italic_u start_POSTSUBSCRIPT italic_N italic_N , italic_y end_POSTSUBSCRIPT - ∇ italic_u start_POSTSUBSCRIPT italic_e italic_x , italic_y end_POSTSUBSCRIPT |. | 2501.10162v1 | 4 |
|
Figure 6.Transport of a constant density from an ellipse domain to a rotated ellipse domain.
Visualization of uniformly sampled source points and their image through exact and approximate transport maps. | 2501.10162v1 | 6 |
|
Figure 7.Transport of a constant density from an ellipse domain to a rotated ellipse domain
Visualization of the transport map∇uNN∇subscript𝑢𝑁𝑁\nabla u_{NN}∇ italic_u start_POSTSUBSCRIPT italic_N italic_N end_POSTSUBSCRIPT. | 2501.10162v1 | 7 |
|
Figure 8.Transport of a constant density from a unit disk domain to an ellipse domain.
Top left: firstx𝑥xitalic_x-coordinate of the transport map∇uNN∇subscript𝑢𝑁𝑁\nabla u_{NN}∇ italic_u start_POSTSUBSCRIPT italic_N italic_N end_POSTSUBSCRIPT;
Top right: Pointwise absolute error|∇uNN,x−∇uex,x|∇subscript𝑢𝑁𝑁𝑥∇subscript𝑢𝑒𝑥𝑥\left|\nabla u_{NN,x}-\nabla u_{ex,x}\right|| ∇ italic_u start_POSTSUBSCRIPT italic_N italic_N , italic_x end_POSTSUBSCRIPT - ∇ italic_u start_POSTSUBSCRIPT italic_e italic_x , italic_x end_POSTSUBSCRIPT |;
Bottom left: secondy𝑦yitalic_y-coordinate of the transport map∇uNN∇subscript𝑢𝑁𝑁\nabla u_{NN}∇ italic_u start_POSTSUBSCRIPT italic_N italic_N end_POSTSUBSCRIPT;
Top right: Pointwise absolute error|∇uNN,y−∇uex,y|∇subscript𝑢𝑁𝑁𝑦∇subscript𝑢𝑒𝑥𝑦\left|\nabla u_{NN,y}-\nabla u_{ex,y}\right|| ∇ italic_u start_POSTSUBSCRIPT italic_N italic_N , italic_y end_POSTSUBSCRIPT - ∇ italic_u start_POSTSUBSCRIPT italic_e italic_x , italic_y end_POSTSUBSCRIPT |. | 2501.10162v1 | 8 |
|
Figure 9.Transport of a Gaussian density to a uniform density.
Left: sampling of points according to Gaussian density (4.1);
Middle: transported solution with exact solution∇uex∇subscript𝑢𝑒𝑥\nabla u_{ex}∇ italic_u start_POSTSUBSCRIPT italic_e italic_x end_POSTSUBSCRIPT;
Right: transported solution with approximated solution∇uNN∇subscript𝑢𝑁𝑁\nabla u_{NN}∇ italic_u start_POSTSUBSCRIPT italic_N italic_N end_POSTSUBSCRIPT.
(N=100000𝑁100000N=100000italic_N = 100000points sampled into the cells of a structured grid of size80×80808080\times 8080 × 80). | 2501.10162v1 | 9 |
|
Figure 10.Transport of a Gaussian density to a uniform density.
Top left: firstx𝑥xitalic_x-coordinate of the transport map∇uNN∇subscript𝑢𝑁𝑁\nabla u_{NN}∇ italic_u start_POSTSUBSCRIPT italic_N italic_N end_POSTSUBSCRIPT;
Top right: Pointwise absolute error|∇uNN,x−∇uex,x|∇subscript𝑢𝑁𝑁𝑥∇subscript𝑢𝑒𝑥𝑥\left|\nabla u_{NN,x}-\nabla u_{ex,x}\right|| ∇ italic_u start_POSTSUBSCRIPT italic_N italic_N , italic_x end_POSTSUBSCRIPT - ∇ italic_u start_POSTSUBSCRIPT italic_e italic_x , italic_x end_POSTSUBSCRIPT |;
Bottom left: secondy𝑦yitalic_y-coordinate of the transport map∇uNN∇subscript𝑢𝑁𝑁\nabla u_{NN}∇ italic_u start_POSTSUBSCRIPT italic_N italic_N end_POSTSUBSCRIPT;
Top right: Pointwise absolute error|∇uNN,y−∇uex,y|∇subscript𝑢𝑁𝑁𝑦∇subscript𝑢𝑒𝑥𝑦\left|\nabla u_{NN,y}-\nabla u_{ex,y}\right|| ∇ italic_u start_POSTSUBSCRIPT italic_N italic_N , italic_y end_POSTSUBSCRIPT - ∇ italic_u start_POSTSUBSCRIPT italic_e italic_x , italic_y end_POSTSUBSCRIPT |. | 2501.10162v1 | 10 |
|
Figure 11.Transport of a Gaussian density to another Gaussian density.
Left: sampling of points according to original Gaussian density;
Right: transported solution with approximated solution∇uNN∇subscript𝑢𝑁𝑁\nabla u_{NN}∇ italic_u start_POSTSUBSCRIPT italic_N italic_N end_POSTSUBSCRIPT.
(N=100000𝑁100000N=100000italic_N = 100000points sampled into the cells of a structured grid of size80×80808080\times 8080 × 80). | 2501.10162v1 | 11 |
|
Figure 12.Transport of a bimodal Gaussian density to a uniform density.
Left: sampling of points according to bimodal Gaussian density (4.2);
Right: transported solution with approximated solution∇uNN∇subscript𝑢𝑁𝑁\nabla u_{NN}∇ italic_u start_POSTSUBSCRIPT italic_N italic_N end_POSTSUBSCRIPT.
(N=100000𝑁100000N=100000italic_N = 100000points sampled into the cells of a structured grid of size80×80808080\times 8080 × 80). | 2501.10162v1 | 12 |
|
Figure 13.Transport of a 3D Gaussian density to a uniform density.
Left: sampling of points inside the unit cube𝒳𝒳\mathcal{X}caligraphic_Xaccording to Gaussian density (4.3);
Right: location of the inside points transported with the approximated solution∇uNN∇subscript𝑢𝑁𝑁\nabla u_{NN}∇ italic_u start_POSTSUBSCRIPT italic_N italic_N end_POSTSUBSCRIPT. | 2501.10162v1 | 13 |
|
Figure 14.Transport of a 3D Gaussian density to a uniform density.
Left: sample of points on the boundary of the unit cube∂𝒳𝒳\partial\mathcal{X}∂ caligraphic_X, uniformly sampled on the boundary
Right: location of the boundary points transported with the approximated solution∇uNN∇subscript𝑢𝑁𝑁\nabla u_{NN}∇ italic_u start_POSTSUBSCRIPT italic_N italic_N end_POSTSUBSCRIPT. | 2501.10162v1 | 14 |
|
Figure 1:The framework of SGMVC. The framework includes: Multi-view Encoder-Decoder Module, which learns feature representations through view-specific encoder-decoder structures. Local Structure Learning Module, reducing false negative samples based on sample neighborhood relationships to enhance instance discriminability. Embedding Structure Learning Module, uncovering latent structural information using Gaussian sampling to minimize embedding discrepancies of the same instance across different views. | 2501.10157v1 | 1 |
|
Figure 4:Visualization of dimensions and their variances related to stereotypical and factual gender signals identified byDual Debiasingalgorithm in 26th layer of Llama 2 13B.
The red dots denote the bias-to-feature thresholdt=0.05𝑡0.05t=0.05italic_t = 0.05.
In2DAMA, the dimension is preserved if stereotypical covariance is below the threshold. | 2501.10150v2 | 8 |
|
Figure 1:Summary of approach. Whilst standard data-driven techniques directly train a time-series model to predict the time-series, this model trains an operational data (current, temperature, voltage, static meta data inc. LIB chemistry and initial capacity) encoder and a simulated curve encoder. Curves are matched to their operational data to enable a zero-shot degradation prediction. | 2501.10492v1 | 1 |
|
Figure 2:Example of an operational degradation curve with the corresponding matched simulated curve. | 2501.10492v1 | 2 |
|
Figure 3:On the left of the image, a degradation curve from the same battery at different points in its lifecycle is shown. In the training process, each of these curves would form a positive and negative pair with the simulated curves on the upper right and lower right, respectively. | 2501.10492v1 | 3 |
|
Figure 4:Future capacity degradation curves as predicted by the model against test LIBs no. 1-4 from 100 cycles of input data. The model has the added benefit against current techniques that it is not limited to a certain output (context) length. | 2501.10492v1 | 4 |
|
Figure 6:Top five most likely degradation paths for test LIB 2 where “Predictionk𝑘kitalic_k” represents thekthsuperscript𝑘thk^{\text{th}}italic_k start_POSTSUPERSCRIPT th end_POSTSUPERSCRIPTmost likely scenario. They all have relatively similar shapes, indicating that the model has learned the degradation behaviour well. | 2501.10492v1 | 6 |
|
Figure 8:A) Zero-shot predictions against test LIB with accelerated degradation & B) Zero-shot predictions against test LIB with standard degradation. Both LIBs taken from dataset from different source to the training set. The model was able to detect whether a particular LIB will experience accelerated degradation, a critical scenario to detect in the operation of LIBs | 2501.10492v1 | 10 |
|
Figure 2:Qualitative chat samples with given multispectral images (images in the figure are contrast enhanced RGB image version of multispectral data just for visualization.). Sample descriptions are given by taking the first a few sentences of model output just to provide visual examples. | 2501.10144v1 | 2 |
|
Figure 3:Comparison of visual features and aligned features projected in a 2D space using t-SNE for EuroSAT miltispectral data. The alignment highlights the transformation of raw visual features into a domain-aligned latent space, showcasing clustering improvements. | 2501.10144v1 | 3 |
|
Figure 4:Test Accuracy Comparison for SpectralGPT and Spectral-LLaVA features through linear probing on EuroSAT dataset with 5-fold cross-validation using various train-test split ratios. | 2501.10144v1 | 4 |
|
Figure 4:100 average training comparison after running 500 episodes. (obtaining a reasonable training score using TD3, TD3+PCA, E-TD3) | 2501.10141v1 | 4 |
|
Figure 1:Miscoverage patterns in standard conformal prediction.Illustration of conditional coverage of standard conformal prediction (standard) over regions of the feature space binned by model confidence and rank of the true class (top) (Conf×RankConfRank\mbox{{{Conf}}}\times\mbox{{{Rank}}}Conf × Rank), and confidence andtrust score(bottom) (Conf×TrustConfTrust\mbox{{{Conf}}}\times\mbox{{{Trust}}}Conf × Trust). We setα=0.1𝛼0.1\alpha=0.1italic_α = 0.1; hence, red bins indicate undercoverage (coverage<<<0.9) and green bins indicate overcoverage (coverage>>>0.9). We split samples into equal-size bins based on rank and trust score. As a special case for ImageNet, we manually define the rank bins, as∼similar-to\sim∼75% test samples have accurate predictions. Models are calibrated using temperature scaling. | 2501.10139v1 | 1 |
|
Figure 2:Average coverage gap between randomly sampled Euclidean balls of fixed radius.We vary the radius on x-axis. Standard errors are reported by error bars. | 2501.10139v1 | 2 |
|
Figure 3:Conditional coverage ofstandard(top) andconditional(bottom) over regions of the feature space binned by model confidence andtrust score(bottom) (Conf×TrustConfTrust\mbox{{{Conf}}}\times\mbox{{{Trust}}}Conf × Trust). | 2501.10139v1 | 3 |
|
Figure 4:Conditional coverage ofstandard(top) andconditional(bottom) over regions of the feature space binned by model confidence and rank of the true class (Conf×RankConfRank\mbox{{{Conf}}}\times\mbox{{{Rank}}}Conf × Rank). | 2501.10139v1 | 4 |
|
Figure 7:Average coverage gap between randomly sampled Euclidean balls of fixed radius.We vary the radius on x-axis. Standard errors are reported by error bars. | 2501.10139v1 | 7 |
|
Figure 1:Examples of the𝐒𝐰𝐝𝐟jsubscript𝐒𝐰𝐝𝐟𝑗\mathbf{Swdf}_{j}bold_Swdf start_POSTSUBSCRIPT italic_j end_POSTSUBSCRIPTof a representative stopword “this” and a topic word “bayesian” | 2501.10137v1 | 1 |
|
Figure 2:The comparison of the original layout with the monotone decreasing layout. Red stands for stopwords while blue stands for topic words. The rectangular shows 64% confidence interval of each dimension and the line in the middle stands for the mean. | 2501.10137v1 | 2 |
|
Figure 4:The scatter plots of the mean and standard deviation of the𝐒𝐰𝐝𝐟jsubscript𝐒𝐰𝐝𝐟𝑗\mathbf{Swdf}_{j}bold_Swdf start_POSTSUBSCRIPT italic_j end_POSTSUBSCRIPTfor topic words (up left), the𝐒𝐰𝐝𝐟jsubscript𝐒𝐰𝐝𝐟𝑗\mathbf{Swdf}_{j}bold_Swdf start_POSTSUBSCRIPT italic_j end_POSTSUBSCRIPTfor stopwords (up right), the𝐒𝐰𝐭𝐟jsubscript𝐒𝐰𝐭𝐟𝑗\mathbf{Swtf}_{j}bold_Swtf start_POSTSUBSCRIPT italic_j end_POSTSUBSCRIPTfor topic words (bottom left) and the𝐒𝐰𝐭𝐟jsubscript𝐒𝐰𝐭𝐟𝑗\mathbf{Swtf}_{j}bold_Swtf start_POSTSUBSCRIPT italic_j end_POSTSUBSCRIPTfor stopwords (bottom right) | 2501.10137v1 | 7 |
|
Figure 1:(a) Simple Function Calling. (b) Complex Function Calling withmulti-step,constraints,parameter value reasoning,long parameter valuesandlong context. Different colors correspond to the corresponding features marked in the figure. | 2501.10132v1 | 1 |
|
Figure 2:Overview of the data collection process. (a) is the high-level process of data collection. (b) is the example of human correction process.(c) is the example of disambiguation process. Thegreypart is removed during annotation. A detailed annotation example is shown in AppendixA.1. | 2501.10132v1 | 2 |
|
Figure 3:Overview ofComplexEval. Different colors represent different API response types. Colorbluerepresents format error with specific error message. Colorgreenrepresents correct function call with corresponding golden API response. Colorredrepresents invalid function call with general error message. | 2501.10132v1 | 3 |
|
Figure 7:An example for golden function call updating. Path on the left is the annotated shortest function call path with three steps. | 2501.10132v1 | 7 |
|
Figure 1:(a) Chest X-rays contain various large (global) and small (local) anatomical patterns, including the right/left lung, heart, spinous processes, clavicle, mainstem bronchus, and the osseous structures of the thorax, which can be utilized for learning global and local embeddings in anatomy. (b) The hierarchical nature of anatomy (eg. The left lung has two lobes, the superior lobex𝑥xitalic_xand the inferior lobey𝑦yitalic_y) calls for anatomical representation with compositionality where the embedding of the whole patch should be the sum of the embeddings of each part. | 2501.10131v1 | 1 |
|
Figure 3:ACE preserves the compositionality of anatomical structures in its learned embedding space. As seen, ACE’s distribution is narrower and taller compared with DINO[6], DropPos[52]and SelfPatch[60], with the mean similarity between embeddings of patches and their compositional parts closer to 1. | 2501.10131v1 | 3 |
|
Figure 4:ACE preserves the decompositionality of anatomical structures in its learned embedding space. As seen, ACE outperforms SSL baselines by a large margin, achieving a high accuracy of 89.01%, which is 30% higher than the second-best baseline. | 2501.10131v1 | 4 |
|
Figure 5:ACE captures semantics-rich features in its learned embedding space. As seen, ACE achieves higher retrieval accuracy compared with other SSL baselines. | 2501.10131v1 | 5 |
|
Figure 6:ACE reflects locality of anatomical structures in its learned embedding space as anemergentproperty. As seen, ACE, unlike the SSL baselines, distinguishes different anatomical structures in its embedding space while keeping identical anatomical structures across patients close to each other. | 2501.10131v1 | 6 |
|
Figure 7:ACE demonstrates its ability to accurately identify anatomical landmarks across different patients. (a) Zero-shot predictions and ground truth of selected query landmarks. (b) Quantitative analysis shows a low prediction error of 61 pixels in10242superscript102421024^{2}1024 start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPTimages, highlighting the robustness of ACE model features in consistent cross-patient anatomical identification. | 2501.10131v1 | 7 |
|
Figure 8:ACE demonstrates superior performance inlimiteddata regimes. As seen in both heart segmentation and pneumothorax classification tasks, ACE surpasses SSL baselines (DINO and POPAR), particularly in few-shot transfer settings. | 2501.10131v1 | 8 |
|
Figure 9:Ablations on the learning objects. Adding the losses improves performance on classification and segmentation. | 2501.10131v1 | 9 |
|
Figure 10:(a) Comparison of proposed ACE with other baselines. AUC scores are shown on the EyePACS diabetic retinopathy classification dataset. (b) ACE’s unsupervised landmark correspondence for paired left and rights fundus image on FIRE dataset. The colored rounds on query and key images are paired ground-truth landmarks, and the red crosses on key images are predicted ones. | 2501.10131v1 | 10 |
|
Figure 11:ACE reflects the symmetry of anatomical structures in its learned embedding space as anemergentproperty. As seen, ACE provides mirrored embeddings for mirrored anatomical structures (e.g., the right and left clavicles, and the right and left rib 5.). | 2501.10131v1 | 11 |
|
Figure 12:ACE demonstrates its ability to boost downstream key point detection tasks. (a) 7 key points are chosen for fine-tuning; (b) the inference detection image where theredpoints are prediction whilegreenpoints are the ground truth situating at the center of the heatmap; (c) comparison between ACE and other pretrained baselines and the lower pixel error the better detection performance. | 2501.10131v1 | 12 |
|
Figure 13:Visualization of Grad-CAM heatmaps. For each column, we provide the heatmap examples for 8 thorax diseases that hold bounding boxes in official labeling. The first row shows the results of our method ACE while the rest of the rows represent the localization of POPAR, DINO, BYOL and Adam. The pink boxes represent the localized ground truth. | 2501.10131v1 | 13 |
|
Figure 1:Comparison between (a, b, c) existing algorithms and (d) our newly proposed MOT tracking framework. | 2501.10129v1 | 1 |
|
Figure 2:In the top row of the image, the individuals represented are extracted from the objects within the MOT17-09 video sequence of the MOT17 dataset. It is worth noting that the girl with a high ponytail reappears multiple times, each time experiencing varying degrees of occlusion. In the second row, the two highlighted objects in the image exhibit a high degree of visual similarity and are located in close proximity to one another. To distinguish between these two objects, we propose utilizing contextual information from the surrounding objects. | 2501.10129v1 | 2 |
|
Figure 1:After perturbing geneTP73, the distribution of geneCENPFchange a lot. However, ground truth collected from comprehensive libraries shows they are independent. | 2501.10124v1 | 1 |
|
Figure 9:Experimental results on a subset of genes with perturbation data. GISL predicts red edges but mismatches with Enrichr(Chen et al.,2013). Black edges are returned by GISL backed by Enrichr. The blue edges indicate that the structure remains undetermined. | 2501.10124v1 | 18 |
|
Figure 10:Examples of distinguishable CI patterns, whereS𝑆Sitalic_Sis the selection variable indicating the selection process,L𝐿Litalic_Lis the latent confounder,X𝑋Xitalic_XandY𝑌Yitalic_Yare observed variables. | 2501.10124v1 | 19 |
|
Figure 11:Examples of DAG-inducing path (non-identifiable cases), whereX𝑋Xitalic_XandY𝑌Yitalic_Yare always d-connected. | 2501.10124v1 | 20 |
|
Figure 12:Examples of causal graphs with three observable variables. The graphs in the dotted box share the same conditional independence relations, and all the other graphs outside the dotted box have different conditional independence relations. | 2501.10124v1 | 21 |
|
Figure 13:Required structure for causal relation, latent confounders, and selection process, Where * means the always d-connected node. | 2501.10124v1 | 22 |
|
Figure 15:The distribution of z-scores for the genes FEV, PTPN1, and SET across all cell lines highlights the variation in sample size. | 2501.10124v1 | 24 |
|
Figure 1:Architecture of PaSa. The system consists of two LLM agents, Crawler and Selector. The Crawler processes the user query and can access papers from the paper queue. It can autonomously invoke the search tool, expand citations, or stop processing of the current paper. All papers collected by the Crawler are appended to the paper queue. The Selector reads each paper in the paper queue to determine whether it meets the criteria specified in the user query. | 2501.10120v1 | 1 |
|
Figure 2:An example of the PaSa workflow. The Crawler runs multiple[Search]using diverse and complementary queries. In addition, the Crawler can evaluate the long-term value of its actions. Notably, it discovers many relevant papers as it explores deeper on the citation network, even when intermediate papers along the path do not align with the user query. | 2501.10120v1 | 2 |
|
Figure 3:Return and value function loss curves during the PPO training process. The smoothing method of the curve in the figures is the exponential moving average(EMA) formula that aligns with the one used in TensorBoard, and the smoothing weight is set to 0.95. | 2501.10120v1 | 3 |
|
Figure 1:Agent infrastructureconsists of technical systems and shared protocols external to agents that are designed to mediate and influence their interactions with and impacts on their environments, including interactions with existing institutions (e.g., legal and economic systems) and actors (e.g., digital service providers, humans, other AI agents). | 2501.10114v1 | 1 |
|
Figure 2:Identity bindinglinks an agent or its actions to an existing legal identity,certificationwould provide assurance about behaviour and properties of an agent instance, and anagent IDswould be a container of information about an agent instance. We illustrateidentity bindingfor a user, but an agent could also be bound to other actors, such as those involved in overseeing it | 2501.10114v1 | 2 |
|
Figure 3:Agent channelsseparate agent traffic from other digital traffic.Oversight layersenable actors to intervene on an agent’s behaviour.Inter-agent communicationhelp to facilitate joint activities amongst groups of agents.Commitment devicesenforce commitments between agents. SeeSection4for discussion. | 2501.10114v1 | 3 |
|
Figure 4:Incident reportingsystems would collect information about and respond to events that could result in harm, whileRollbacksare tools that would void or undo an agent’s actions. SeeSection5for further discussion. | 2501.10114v1 | 4 |
|
Figure 2:Overview of our proposed DiffVSR framework. (a) The overall model architecture integrates enhanced UNet and VAE decoder for high-quality frame restoration. (b) Detailed designs of our modified UNet and VAE decoder blocks for better feature extraction and reconstruction. (c) Progressive Learning Strategy that enables stable training and robust performance across various degradation levels. (d) Multi-Scale Temporal Attention (MSTA) mechanism designed for capturing temporal dependencies at different scales. Notably, spatial layer includes ResBlock2D and Spatial Attention, while temporal layer contains ResBlock3D, Temporal Attention and MSTA module. | 2501.10110v2 | 2 |
|
Figure 3:Illustration of our temporal consistency enhancement approach. Noise rescheduling mechanism that reuses and reorders noise frames across overlapping regions. Interweaved Latent Transition (ILT) scheme that performs position-based latent interpolation between adjacent subsequences. This lightweight post-processing solution effectively ensures temporal consistency without additional training or computational overhead. | 2501.10110v2 | 3 |
|
Figure 4:Qualitative comparisons on synthetic low-quality videos from YouHQ40 and UDM10 datasets. Among the tested methods, only our approach can recover the accurate facial beard texture and produce detailed text on the billboard.(Zoom-in for best view) | 2501.10110v2 | 4 |
|
Figure 5:Qualitative comparisons on real-world test videos from MVSR4x and RealVideo10 datasets. Our method effectively leverages the advantages in generating high-quality results. When compared to existing methods, it notably excels in its restoration capabilities, successfully recovering architectural textures and depth details. In particular, our method accurately restores the natural skin texture of sea turtles in underwater scenes, while other methods incorrectly render the turtle’s skin with stone-like textures.(Zoom-in for best view) | 2501.10110v2 | 5 |
|
Figure 6:Temporal profiles on a RealVideo10 sequence. By stacking frame scan lines (yellow dashed line) chronologically, we observe that image-based method exhibits temporal flickering, while our approach demonstrates both superior restoration quality and smooth temporal transitions.(Best viewed on digital displays) | 2501.10110v2 | 6 |
|
Figure 7:Quantitative comparison of user preferences among different video restoration methods. Our method achieves the highest preference rate (82.0%) in the user study, significantly outperforming RealBasicVSR[9], MGLD[48], and UAV[53]. | 2501.10110v2 | 7 |
|
Figure 8:Illustration of three training strategy variants: Direct Learning Strategy, 2-Stage Learning Strategy, and Progressive Learning Strategy. The proposed Progressive Learning Strategy can not only stabilize the training procedure but also improve performance. | 2501.10110v2 | 8 |
|
Figure 10:Analysis of the trade-off between video quality (MUSIQ) and temporal consistency (Warping Error) on YouHQ40 dataset. Methods towards the lower-right corner achieve better overall performance, with higher MUSIQ and lower warping error. | 2501.10110v2 | 10 |
|
Figure 12:Qualitative comparisons on synthetic datasets. Our method demonstrates superior capability in recovering accurate facial details and textual information, while other methods struggle with either over-smoothing or detail distortion.(Zoom-in for best view) | 2501.10110v2 | 12 |
|
Figure 1:High level architectural design of our experiments.
The static + historic series decoder block MLP (MQCNN) is in fact a complex collection of MLPs for different forecasting horizons and quantiles.
Pre-trained LLMs considered are GPT-2, Flan-T5, and MPT-7B. | 2501.06386v1 | 1 |
|
Figure 2:Layer-level weight analysis identifies sub-optimal model architecture and predicts forecasting accuracy.Complementary cumulative distribution function (CCDF) of weight matrix spectrum for the first embedding layer (left) and the output layer (middle). The rightmost plot shows evolution of P50 quantile test loss across epochs for both architectures, where markers and lines are colored according to the fittedα𝛼\alphaitalic_α. | 2501.06386v1 | 2 |
|
Figure 4:HTSR metrics predict forecasting accuracy across architectures (varying base FPTs) and within architecture across epochs.P50 quantile test loss by epoch for GPT-2- and Flan-T5-based architectures. Markers and lines are colored according to theα𝛼\alphaitalic_αmetric (left) and the stable rank metric (right). Within the same architecture over different epochs, a higher metric value generally results in a higher accuracy; at a fixed epoch between different architectures, a higher metric value generally results in a higher accuracy. | 2501.06386v1 | 13 |
|
Figure 6:HTSR metrics predict forecasting accuracy across architectures (varying input/output layers) and within architecture across epochs.P50 quantile test loss by epoch for architectures with a Linear and MLP embedding/decoding layer, respectively. Markers and lines are colored according to theα𝛼\alphaitalic_αmetric (left) and the stable rank metric (right). | 2501.06386v1 | 26 |
|
Figure 7:ESD of the embedding layer (top) and the output layer (bottom), when FPT is included (left) and not included (middle) in the architecture. For this figure, Flan-T5-small is used as the FPT, and we analyze architectures that adopt linear input and output layers. Red vertical lines correspond to theλminsubscript𝜆\lambda_{\min}italic_λ start_POSTSUBSCRIPT roman_min end_POSTSUBSCRIPTparameter of the PL distribution in (3) chosen by minimizing the KS distance, and red dashed lines represent the PL distribution for the fittedα𝛼\alphaitalic_α. Additionally, the tail of the complementary cumulative distribution function (CCDF) is shown for each layer (right). | 2501.06386v1 | 29 |
Subsets and Splits