image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
Figure 15:Output of the1111-layer residual MLP target model compared to true labels for the full set of100100100100one-hot inputs.Left: Output at all output indices over the set of inputs. The point color indicates the active input feature, and label values are in red.Right: Output at indexi𝑖iitalic_ifor inputs withxi∈[0,1]subscript𝑥𝑖01x_{i}\in[0,1]italic_x start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT ∈ [ 0 , 1 ]andxj=0subscript𝑥𝑗0x_{j}=0italic_x start_POSTSUBSCRIPT italic_j end_POSTSUBSCRIPT = 0forj≠i𝑗𝑖j\neq iitalic_j ≠ italic_i. Line colors indicate the input feature index. | 2501.14926v3 | 15 |
|
Figure 16:MSE for APD trained with batch top-k=1.28𝑘1.28k=1.28italic_k = 1.28in the1111-layer residual MLP setting for samples with a single active input feature (i.e. one-hot), averaged over100100100100k samples.Top:Comparison of the target model with the APD model when activating exactly one parameter component in each sample (i.e. top-k=1𝑘1k=1italic_k = 1).Bottom:Comparison of the target model with the APD model using batch top-k=1.28𝑘1.28k=1.28italic_k = 1.28. The batch top-k𝑘kitalic_kmask is applied to the original training distribution and then samples without exactly one active input feature are filtered out. | 2501.14926v3 | 16 |
|
Figure 17:MSE for APD trained with batch top-k=1𝑘1k=1italic_k = 1in the1111-layer residual MLP setting for samples with a single active input feature (i.e. one-hot), averaged over100100100100k samples.Top:Comparison of the target model with the APD model when activating exactly one parameter component in each sample (i.e. top-k=1𝑘1k=1italic_k = 1).Bottom:Comparison of the target model with the APD model using batch top-k=1𝑘1k=1italic_k = 1. The batch top-k𝑘kitalic_kmask is applied to the original training distribution and then samples without exactly one active input feature are filtered out. | 2501.14926v3 | 17 |
|
Figure 18:Relationship in the1111-layer residual MLP setting between: (y-axis) the average number of active APD parameter components when using batch top-k=1.28𝑘1.28k=1.28italic_k = 1.28, and (x-axis) the MSE between the target model outputs and the APD model when activating exactly one parameter component in each sample (i.e. top-k=1𝑘1k=1italic_k = 1). MSE is measured only on samples with a single active input feature. | 2501.14926v3 | 18 |
|
Figure 19:Output of the2222-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 | 19 |
|
Figure 20:Output of the2222-layer residual MLP target model compared to true labels for the full set of100100100100one-hot inputs.Left: Output at all output indices over the set of inputs. The point color indicates the active input feature, and label values are in red.Right: Output at indexi𝑖iitalic_ifor inputs withxi∈[0,1]subscript𝑥𝑖01x_{i}\in[0,1]italic_x start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT ∈ [ 0 , 1 ]andxj=0subscript𝑥𝑗0x_{j}=0italic_x start_POSTSUBSCRIPT italic_j end_POSTSUBSCRIPT = 0forj≠i𝑗𝑖j\neq iitalic_j ≠ italic_i. Line colors indicate the input feature index. | 2501.14926v3 | 20 |
|
Figure 21:MSE for APD trained with batch top-k=1.28𝑘1.28k=1.28italic_k = 1.28in the2222-layer residual MLP setting for samples with a single active input feature (i.e. one-hot), averaged over100100100100k samples.Top:Comparison of the target model with the APD model when activating exactly one parameter component in each sample (i.e. top-k=1𝑘1k=1italic_k = 1).Bottom:Comparison of the target model with the APD model using batch top-k=1.28𝑘1.28k=1.28italic_k = 1.28. The batch top-k𝑘kitalic_kmask is applied to the original training distribution and then samples without exactly one active input feature are filtered out. | 2501.14926v3 | 21 |
|
Figure 22:MSE for APD trained with batch top-k=1𝑘1k=1italic_k = 1in the2222-layer residual MLP setting for samples with a single active input feature (i.e. one-hot), averaged over100100100100k samples.Top:Comparison of the target model with the APD model when activating exactly one parameter component in each sample (i.e. top-k=1𝑘1k=1italic_k = 1).Bottom:Comparison of the target model with the APD model using batch top-k=1𝑘1k=1italic_k = 1. The batch top-k𝑘kitalic_kmask is applied to the original training distribution and then samples without exactly one active input feature are filtered out. | 2501.14926v3 | 22 |
|
Figure 23:Relationship in the2222-layer residual MLP setting between: (y-axis) the average number of active APD parameter components when using batch top-k=1.28𝑘1.28k=1.28italic_k = 1.28, and (x-axis) the MSE between the target model outputs and the APD model when activating exactly one parameter component in each sample (i.e. top-k=1𝑘1k=1italic_k = 1). MSE is measured only on samples with a single active input feature. | 2501.14926v3 | 23 |
|
Figure 24:Output of multiple2222-layer residual MLP 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)). | 2501.14926v3 | 24 |
|
Figure 25:MSE losses of the2222-layer residual MLP 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 which uses one monosemantic neuron per input feature. | 2501.14926v3 | 25 |
|
Figure 1:Overview of the proposed silhouette-based differentiable rendering method for 3D/2D angiographic registration using two views. The framework optimizes the alignment of 2D DSA images with the 3D vascular model.Click the image to play the video in a browser. | 2501.14918v1 | 1 |
|
Figure 2:First (top) and final (bottom) iterations of the registration process using the two AP and LAT images for synthetic data.Click the image to play the video in a browser. | 2501.14918v1 | 2 |
|
Figure 1:We introducing a self-reflection method inspired by the Hegelian dialectic to generate innovative ideas. We examine with two temperature configurations: a dynamic annealing process that encourages high creativity initially and gradually decreases as the model converges on its final idea, and a fixed temperature setting to ensure consistent behavior. We utilize a Multi Agent Majority Voting (MAMV) paradigm to evaluate the novelty and validity of the generated synthesis. | 2501.14917v3 | 1 |
|
Figure 1:Processing speed vs. accuracy for various SfM methods.Our work significantly decreases the runtime across various sizes of image collections compared to traditional pipelines while obtaining comparable accuracy. Results are measured on the Tanks&Temples dataset. | 2501.14914v1 | 1 |
|
Figure 2:Light3R-SfM Pipeline.Given an unordered set of images, we first encode them to obtain image tokens from which we average pool global features for constructing a shortest path tree. We next feed image tokens into our attention-based latent global alignment to enable global context sharing. Afterwards, for each edge in the SPT, we decode pairwise pointmaps using the implicitly aligned feature tokens. Finally, we use global accumulation to obtain globally aligned pointmaps per image. | 2501.14914v1 | 2 |
|
Figure 3:Qualitative comparison on a Waymo scene.Note how the MASt3R-SfM reconstruction does not truthfully reconstruct the 90° turn, while Spann3R predictions degrade after tens of frames. | 2501.14914v1 | 3 |
|
Figure 4:Reconstructing opposite-oriented cameras.After conditioning Light3R-SfM’s decoder with the output from our global latent alignment, it is able to predict pointmaps even for images recorded in opposite directions, suggesting the latent global alignment has learned a representation of the entire scene. | 2501.14914v1 | 4 |
|
Figure 6:Global confidence map (right) produced by Light3R-SfM for an image of a sequence containing dynamic objects (left). | 2501.14914v1 | 6 |
|
Figure 1:Comparison of captioning methods:Rule-based captions are limited in detail. Unimodal LLM captions are fluid but often generic. Wide-area scenes covering diverse structures and objects require semantically rich descriptions. We leverage the semantic density of maps to generate comprehensive and detailed captions. | 2501.14905v1 | 1 |
|
Figure 3:A sample from the fMoW-mm dataset. The generated caption accurately incorporates information from the satellite image, map, and metadata. | 2501.14905v1 | 3 |
|
Figure 4:Ablations. (a) Map Resolution:Higher resolution reduces hallucination rates and uncertainty in generated captions.(b) Map Types:Using landmarks-only gives the best balance, reducing hallucinations while limiting uncertainty.(c) Prompt Ensembling:Combining captions from multiple prompts did not significantly impact the metrics, however increasing from 3 to 5 prompts may result in repeated hallucinations that propagate into the final caption. | 2501.14905v1 | 4 |
|
Figure 1:Category-level 3D reconstruction: on the top are input images from NOCS dataset[4]and on the bottom are predicted 3D shapes; on the left are examples of training instances and on the right are testing instances. Our model predicts complete 3D shapes of instances unknown during training based on their RGB appearance and category shape prior learned from data. | 2501.14896v1 | 1 |
|
Figure 2:Overview of Glissando-Net: the RGB image is processed by an encoder-decoder; the feature maps from the RGB decoder are then transformed and concatenated with features in the point cloud decoder, which takes in a latent code𝐳𝐳\mathbf{z}bold_z; the outputs from the point-cloud decoder are processed by a fully-connect layer to regress the 3D shape in canonical pose and two independent multilayer perceptrons (MLP) to estimate its pose; in training, the ground-truth canonical shape is an additional input to a point cloud encoder to learn the distribution of𝐳𝐳\mathbf{z}bold_zand feature maps from the RGB decoder are transformed to augment point cloud features. | 2501.14896v1 | 2 |
|
Figure 3:Details of RGB and point cloud encoder-decoder from Glissando-Net. Red arrows indicate skip connections between RGB encoder and decoder layers. Features 1–4 from the RGB decoder are fed into the point cloud encoder-decoder. The point cloud encoder uses set abstraction (SA) layers with sampled points from PointNet++, which are projected back to the RGB feature map using RoI align. Features 1–4 are concatenated with the point cloud decoder outputs via 2D convolution and fully connected layers. | 2501.14896v1 | 3 |
|
Figure 7:Feature Map Visualization for 2D and 3D. (a) shows the input RGB image. (b) shows the 2D image feature map. (c) shows the point cloud generated without the RGB encoder and decoder. (d) shows the point cloud generated with the RGB encoder and decoder. (e) shows the ground truth 3D shape. | 2501.14896v1 | 7 |
|
Figure 8:Qualitative Results on Objectron Dataset. The first row shows the input image. The second row shows the reconstructed 3D shape. | 2501.14896v1 | 8 |
|
Figure 9:Qualitative results onsyntheticNOCS data. Top to bottom: input image, ground-truth 3D shape with ground-truth rotation, and our predicted 3D shape with predicted rotation. | 2501.14896v1 | 9 |
|
Figure 10:Qualitative results onrealNOCS data. Top to bottom: input image, ground-truth 3D shape with ground-truth rotation, and our predicted 3D shape with predicted rotation. | 2501.14896v1 | 10 |
|
Figure 11:Overlay results on NOCS data. Top to bottom: input image, project the predicted point cloud to the input image. | 2501.14896v1 | 11 |
|
Figure 12:Some examples where our algorithm fails. The two examples on the left are form the bottle category, while the example on the right is from the mug category with occlusion from the top. The left two examples are of very unusual shapes. The rightmost example is occluded at the top. | 2501.14896v1 | 12 |
|
Figure 13:Qualitative results on Pix3D. Top to bottom: input image, ground-truth 3D shape with ground-truth rotation, and our predicted 3D shape with predicted rotation.Notice that many of the shapes are partially occluded in the images. | 2501.14896v1 | 13 |
|
Figure 2:The reliability plot comparing the predicted cumulative probability and empirical cumulative probability obtained from observing real data. The red dashed line depicts a perfectly calibrated model whose prediction exactly matches the observation results. Exiting uncertainty forecasters are usually miscalibrated in creating curves far from the ideal calibration line. The miscalibration leads to erroneous and off-centered CI estimation. | 2501.14894v1 | 2 |
|
Figure 3:The heteroskedastic regression model used to perform the probabilistic regression. Inputs, x, into the model are composed of left eye image, right eye image, and head angles, the eye images are processed by ResNet with the same feature extraction structure but different weights. The head angles are concatenated to outputs from the linear layer to be considered for downstream inferences. The models’ outputs are two Gaussian distributions for pitch and yaw gaze angles, described by mean angle values and their associated variances. Variances are used to assess prediction uncertainties. | 2501.14894v1 | 3 |
|
Figure 4:Reliability plot comparing the behaviors of data used for calibration (black curve), data used for testing (blue curve), post-calibration data (cyan curve), and the ideal calibrated line (red dashed line). In same-subject calibration cases, bias behaviors in calibration and testing are alike, leading to close-to-ideal calibration (top row). In the cross-subject calibration cases, bias behaviors could be different, leading to unideal calibration (bottom row). | 2501.14894v1 | 4 |
|
Figure 5:Errors evaluating the difference between desired CI and actual true value inclusion rate from all testing cases. The two baseline, uncalibrated models have the highest level of error magnitude and variations, which makes them unsuitable for providing reliable uncertainties for HMI applications. The same-subject calibration achieved the best and most consistent error performance, which makes it the ideal application configuration. The below-zero portions in the cross-subject calibration showed that this calibration setup is likely lead to underconfidence. Overall, the training dataset does not have strong influences on error distribution. | 2501.14894v1 | 5 |
|
Figure 1:Overall framework in three stages:
(1) building/updating the causal graph,
(2) chain-of-thought–driven retrieval,
(3) multi-stage path processing. | 2501.14892v1 | 1 |
|
Figure 2:An example of chain-of-thought–driven retrieval.
Each CoT segment triggers a partial query to the causal subgraph,
producing candidate paths that are later fused. | 2501.14892v1 | 2 |
|
Figure 3:Path scoring and pruning.
Each path is evaluated based on CUI overlap, semantic overlap, and length heuristics. | 2501.14892v1 | 3 |
|
Figure 1:Selecting an AutoAIS evaluator based solely on balanced accuracy (BAcc) hides several underlying inconsistencies. Consider gpt-4-turbo and Bespoke-7B with comparable BAcc onLLM-AggreFact. The two evaluators have (a) low instance-level labeling consistency and (b) different true positive and true negative error rate trade-offs. (c) This results in different system-level evaluations when the evaluators are used downstream to evaluate the factuality of NLG systems. In several cases, one evaluator underestimates the human-labeled error rate while the other overestimates it. | 2501.14883v2 | 1 |
|
Figure 1:Each image in our dataset is a concatenation of a math problem on the left with a student response on the right. Teachers describe the student’s response to the problem, and then a model, such as GPT-4o shown here, writes QA pairs extracted from facets of the description. More example images, along with teacher-written QA, are shown in Figure3. | 2501.14877v1 | 1 |
|
Figure 2:For some annotators, their recorded descriptions of images are longer or require less time than typed ones. Annotation length is calculated based on white-spaced-separated tokens. | 2501.14877v1 | 2 |
|
Figure 3:Examples of teachers’ answers to a question asking about possible errors in students’ responses to math problems. All three examples of students’ hand-drawn responses are for the same math problem asking students to draw and shade units on fraction strips to show 4 thirds, shown on the left. | 2501.14877v1 | 3 |
|
Figure 4:A screenshot of our recording website, where teachers would view an image from our dataset and either write or record a description of the student’s response. Typically, “unknown teacher ID” would include the currently annotating teacher’s ID. | 2501.14877v1 | 4 |
|
Figure 5:A screenshot of the interface teachers used to write answers to teacher-written questions about students’ responses. Typically, “unknown teacher ID” would include the currently annotating teacher’s ID. | 2501.14877v1 | 5 |
|
Figure 1:(a) Previous driving world models focus solely on generation, which predicts scene evolution. (b) Large language models for driving are limited to scene understanding. (c) A straightforward unification manner by utilizing the future generator and large language model separately. (d) The proposed simple framework unifies the understanding of the surrounding 3D scene and generates scene evolution. | 2501.14729v1 | 1 |
|
Figure 1.Relightable Full Body Gaussian Codec Avatars.We present
the first approach that enables reconstruction, relighting and expressive
animation of full-body avatars including body, face, and hands. Our
approach combines learned, orientation-dependent diffuse radiance transport
and deferred-shading-based specular radiance transport to enable complex
light transport such as global illumination for fully articulated human
bodies. | 2501.14726v1 | 1 |
|
Figure 6.Capture Dome.Our multi-camera light stage with 512 cameras and 1024 controllable light sources. The dome has a radius of2.752.752.752.75meters. Each camera has 24 mega-pixels resolution and records video with up to 90Hz. | 2501.14726v1 | 15 |
|
Figure 9.Relighting result on unseen motion.We show environment-map-based relighting on the left two columns and point-light-based relighting on the right two columns. | 2501.14726v1 | 24 |
|
Figure 1:Overview of the CodeMonkeys system.Left:We retrieve codebase context by using models to first identify relevant files and then rank them relative to each other.Middle:We generate a codebase edit and testing script using a pair of multi-turn state machines that iterate based on execution feedback. We run these state machines multiple times in parallel to generate 10 edits and tests for every issue.Right:We select between candidate edits by identifying the candidates that pass the most generated tests and asking a model to decide between these top candidates. For details about our system’s three state machines, see Figure4. | 2501.14723v2 | 1 |
|
Figure 2:Measuring CodeMonkeys performance across the three subtasks we identify in Section2(context, generation, and selection). Note that modifying the approach to one subtask can influence the performance on other subtasks as well. For example, generating more candidate edits could increase coverage but make selection harder. | 2501.14723v2 | 2 |
|
Figure 3:Left: Measuring recall (the fraction of SWE-bench problems with context windows that contain all needed files) as we increase the context window size limit.
With the 128k token limit that we use for later experiments, 92.6% of instances have the correct files in context.Right: Visualizing the distribution of context compression factors across SWE-bench problems, i.e. the ratio between the cumulative token count of files scanned by the relevance model and the cumulative token count of files we include after relevance + ranking. | 2501.14723v2 | 3 |
|
Figure 5:Measuring coverage (left) and score when using majority-voting selection (right) as we sweep over the number of serial iterations per editing state machine and the number of parallel state machines sampled per problem. Each colored curve corresponds to a different number of parallel state machines, and the dots along each curve correspond to increased numbers of sequential iterations per state machine. The first few serial iterations have a large impact on improving performance. However, past that point, different configurations with similar costs lead to similar performance, particularly for coverage. | 2501.14723v2 | 5 |
|
Figure 6:Comparing our selection methods when applied to the candidate edits generated by the CodeMonkeys editing state machines. Our best performing selection method – the selection state machine after top-3 filtering with generated tests – recovers approximately half of the difference between the random selection floor and the oracle selection ceiling (i.e. coverage). | 2501.14723v2 | 6 |
|
Figure 6:Comparing our selection methods when applied to the candidate edits generated by the CodeMonkeys editing state machines. Our best performing selection method – the selection state machine after top-3 filtering with generated tests – recovers approximately half of the difference between the random selection floor and the oracle selection ceiling (i.e. coverage). | 2501.14723v2 | 7 |
|
Figure 7:Left:Comparing the impact of scaling the number of parallel samples and sequential iterations on majority voting score between Claude and DeepSeek-V3. Each line corresponds to a fixed amount of samples, with each dot on the line being a different maximum number of sequential iterations. We see that although Claude can achieve a higher overall score, DeepSeek-V3 can achieve86.8%percent86.886.8\%86.8 %percent of the score at a fraction of the cost.Center:A more granular view of the majority voting scaling for DeepSeek-V3.Right:Coverage for DeepSeek-V3 as a function of the number of parallel samples and sequential iterations. We highlight that coverage is continuing to scale with increased inference compute. | 2501.14723v2 | 9 |
|
Figure 1:The proposed framework to automatically measure the consistency of health-related answers in EN, ZH, TR and DE. | 2501.14719v1 | 1 |
|
Figure 1:The proposed framework to automatically measure the consistency of health-related answers in EN, ZH, TR and DE. | 2501.14719v1 | 2 |
|
Figure 3:The answers to the question "Does discontinuation of denosumab increase the risk of vertebral fractures?": the EN answer on the left and the TR answer on the right. Both were parsed by GPT-4o using the proposed discourse ontology. | 2501.14719v1 | 4 |
|
Figure 1:FlexiGPT is used for two settings: (1) pruning a model to reduce parameters with minimal performance costor(2) extending a model to increase performance with minimal parameter cost.Left:For pruning models (setting 1), we prune entire blocks and replace them using weight sharing and learned adapters.Right:For extending models (setting 2), we repeat block patterns in the model using weight sharing and learned adapters. | 2501.14713v2 | 1 |
|
Figure 3:Overview of the FlexiGPT pruning process.Left:We prune model blocks with the lowest scores based on (1).Center:We select replacement blocks with high similarity using (5).Right:We add feature normalization and learn adapters to recover performance. | 2501.14713v2 | 6 |
|
Figure 5.Distribution ofXAsubscript𝑋𝐴X_{A}italic_X start_POSTSUBSCRIPT italic_A end_POSTSUBSCRIPTin simulated real, FiND, adapted and warped world per protecteda𝑎aitalic_aand unprotected groupa′superscript𝑎′a^{\prime}italic_a start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT | 2501.14710v1 | 12 |
|
Figure 6.Distribution ofXAsubscript𝑋𝐴X_{A}italic_X start_POSTSUBSCRIPT italic_A end_POSTSUBSCRIPTon real-world, adapted and warped HMDA data per protecteda𝑎aitalic_aand unprotected groupa′superscript𝑎′a^{\prime}italic_a start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT | 2501.14710v1 | 13 |
|
Figure 1.The proposed framework starts by pre-training the system dynamics model on historical data. Then, the pre-trained model parameters are fed into the convex optimization control policy where they keep being updated on representative scenarios to minimize the loss metric evaluating the mismatch between the expected and observed system states. | 2501.14708v1 | 1 |
|
Figure 2.The building is made of three similar floors. Each floor is equipped with its own variable air volume air handling unit responsible for the thermal comfort of the five zones. | 2501.14708v1 | 2 |
|
Figure 3.Ambient temperature profiles of the ten medoids. The three first labels are extreme scenarios. | 2501.14708v1 | 3 |
|
Figure 5.The buildingRCmodel is pre-trained on one year of historical data before a small noise is applied on the parameters. They are then updated to perform the best day-aheadHVACscheduling over 10 representative days with respect to the task-aware MAE loss. | 2501.14708v1 | 5 |
|
Figure 6.(a) TheDFLsupervised loss (hierarchical weighted MAE) converges at much lower levels than initially; (b) the error mean converges towards zero and the standard deviation reduces; (c) the expected cost becomes a more accurate prediction of the ex-post cost while diminishing. | 2501.14708v1 | 6 |
|
Figure 8.Scheduling for the coldest day obtained by theDFLRCmodel at epoch 34. Subfigures are (a) the zonal indoor temperature of the top floor; (b) the zonalHVACelectric power of the top floor, and (c) the aggregatedHVACpower for the whole building and each floor. The blue shades indicate the temperature deviations causing an objective value penalty lower than a given threshold. Temperature control is more loose during the night and evening because of the worker absence. The mismatch between the expected and ex-postHVACpowers in (b) can be explained by the intrinsicRCmodel limitations and the task-aware loss which gives little importance to accurate zonal predictions. | 2501.14708v1 | 11 |
|
Figure 9.Ambient temperature distribution of each data set. Whiskers represent minimum and maximum temperatures. | 2501.14708v1 | 12 |
|
Figure 2:The distribution of lengths (i.e., number of characters) of reductions in the dataset. Most reductions have lengths between 1000 and 3000 characters. The minimum is 939, the maximum is 5789, and the mean is 2180. | 2501.14705v1 | 2 |
|
Figure 4:Two-dimensional model resembling a slice of a human head. (a) Reference model containing five different compartments of importance for electrophysiology of the human head: grey matter, white matter, cerebrospinal fluid, skull and other soft tissues representing the scalp. (b) Reference model included in a elliptic shape. The dark blue region can be interpreted as a fixed helmet of constant conductivity with electrodes attached to its exterior. (c) The head anatomy inside the helmet of one simulated patient. | 2501.14704v2 | 4 |
|
Figure 5:Samples of the training and test datasets. (a) The stroke is represented by circular inclusion. This type of inclusion is considered for the training dataset and for the first set of the test dataset. (b) A conductivity sample from the second test dataset where the stroke is modeled as an elliptic inclusion. (c) A conductivity sample from the third test dataset where the stroke is modeled as multiple inclusions containing one disc and one elliptic inclusion. | 2501.14704v2 | 5 |
|
Figure 2:Guidelines for referrals forvenous echocolordoppler of the lower limbsin the Lombardy Region. Translated from the Appendix 3 of[26]. | 2501.14701v1 | 2 |
|
Figure 4:Confusion matrix of the clustering assignment to the labels on the manually annotated test set, computed considering the number of clinical questions (A) and the number of referrals (B). | 2501.14701v1 | 4 |
|
Figure 5:Examples of clustered clinical questions. English translations do not completely reflect the abbreviations and typos of the original Italian texts. | 2501.14701v1 | 5 |
|
Figure 6:Prescription appropriateness grouped by LHA (A), type of prescribing physicians (B) and year (C) | 2501.14701v1 | 6 |
|
Figure D.1:Prescription appropriateness forflexible endoscope colonscopygrouped by LHA (A), type of prescribing physicians (B) and year (C) | 2501.14701v1 | 7 |
|
Figure 2:Expected connections from the CAGE 2 layout configuration.
Connections among a subnet are unrestricted while connections between subnets are limited to specific hosts. | 2501.14700v2 | 2 |
|
Figure 4:Examples of graph encoded observations. Edge labels show the number of connections and node labels show the open ports per host (more features are available per node, see figure3). | 2501.14700v2 | 4 |
|
Figure 5:Probability distribution over global actions and actions per node, as predicted by the GAT policy for a particular state.
The probabilities are derived through the softmax transformation of raw scores.
Each node (rows) outputs the scores of 11 local actions (columns) and 2 global actions (not displayed).
The 2 global action scores are the sum of the corresponding global action scores per node, allowing a trained GAT policy to generalise to graphs of different size. | 2501.14700v2 | 8 |
|
Figure 6:Distribution of reward-to-go values per episode step. Higher peaks in the trained policy distribution near zero reward showcase improved performance over random policies. Dips in reward, appearing as leftward modes in the distributions, correspond to episodes experiencingoperational server impact(as detailed in Table1). | 2501.14700v2 | 9 |
|
Figure 8:Unexpected connections observed.
This highlights a limitation of the structural inductive bias of GNNs when the expected layouts do not accurately reflect the input distribution, potentially leading to overfitting[31].
However, GNNs are still able to handle these topological deviations, unlike other architectures. | 2501.14700v2 | 11 |
|
Figure 1:GraPPI for target identification (Target ID). Based on the two inputs- an initial protein and therapeutic impact on the initial protein- GraPPI recommends PPI pathways with explanations and retrieves text based on previous work. | 2501.16382v1 | 1 |
|
Figure 2:Overview ofGraPPIframework. The input of the users toGraPPIis the name of the initial protein and the therapeutic impact query. The outputs are recommended PPIs with AI-generated explanations and retrieved information from the database. | 2501.16382v1 | 2 |
|
Figure 3:Box plots of results under different graph sizes. The blue color in the background indicates the level of difference between the two groups. Darker blue represents a more significant difference. For plots (a), (b), and (c), the red and orange boxes represent different accuracies ofGraPPIand the system directly using protein annotations, respectively. For plot (d), the green and purple boxes indicate the number of input tokens they have.Rawrefers to the methods using raw annotations texts as contexts whileOursutilizes the edge explanations with more concise representation of biomedical context. | 2501.16382v1 | 3 |
|
Figure 4:Part of the results of the case study showing the input and output content regarding certain recommended PPI signaling pathways. | 2501.16382v1 | 4 |
|
Figure 1:Large performance variations (here measured by AUC) over different hyperparameter configurations for ANEMONE[25]on various benchmark datasets. Using labeled data and only reporting the best possible performance leads to severe overestimation of model performance. For instance, the green squares on Cora, CiteSeer, and PubMed are reported by[25](the other datasets were not used). Similar results are observed for other algorithms (see AppendixBfor details). The red triangles represent the results obtained by our internal evaluation strategy, showing its potential for automating truly unsupervised anomaly detection. | 2501.14694v1 | 1 |
|
Figure 2:Self-supervised learning based graph anomaly detection methods can be subdivided intogenerative basedmethods andcontrastive basedmethods. Agenerative basedmethod generally involvesgraph structure reconstructionandnode attributes reconstruction. Acontrastive basedmethod usually consists of agraph augmentation moduleand acontrastive learning module. | 2501.14694v1 | 2 |
|
Figure 3:Sensitivity analysis ofk𝑘kitalic_k(for our proposed AutoGAD) on dataset CiteSeer with all investigated SSL-based GAD algorithms. It can be seen that AutoGAD remains stable as long ask𝑘kitalic_kis not drastically distant from the actual anomaly ratio (namely4.5%percent4.54.5\%4.5 %) all for SSL-based GAD algorithms. | 2501.14694v1 | 3 |
|
Figure 4:Performance of AutoGAD across different granularity levels of search grids using ANEMONE on the Cora, ACM, and Facebook datasets. Similar trends were observed for other anomaly detectors and datasets, which are omitted for brevity. | 2501.14694v1 | 4 |
|
Figure 5:Performance variations over different HP configurations for CoLA[47]on different benchmark datasets. | 2501.14694v1 | 5 |
|
Figure 6:Performance variations over different HP configurations for GRADATE[13]on different benchmark datasets. | 2501.14694v1 | 6 |
|
Figure 7:Performance variations over different HP configurations for SL-GAD[102]on different benchmark datasets. | 2501.14694v1 | 7 |
|
Figure 8:Performance variations over different HP configurations for Sub-CR[91]on different benchmark datasets. | 2501.14694v1 | 8 |
|
Figure 9:Performance variations over different HP configurations for CONAD[80]on different benchmark datasets. | 2501.14694v1 | 9 |
|
Figure 10:Performance variations over different HP configurations for DOMINANT[10]on different benchmark datasets. | 2501.14694v1 | 10 |
|
Figure 11:Performance variations over different HP configurations for AnomalyDAE[15]on different benchmark datasets. | 2501.14694v1 | 11 |
|
Figure 12:Performance variations over different HP configurations for GUIDE[86]on different benchmark datasets. | 2501.14694v1 | 12 |
|
Figure 13:Performance variations over different HP configurations for GAAN[7]on different benchmark datasets. | 2501.14694v1 | 13 |
Subsets and Splits