image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
Figure 2:System Diagram forXAttnMark.XAttnMarkconsists of a watermark generator and a watermark detector, with a shared embedding table that facilitates message decoding through a cross-attention module. In the generator part, we first employ an encoder network to encode the audio latent and then apply a temporal modulation to hide the message. The modulated latent is then fed into a decoder to produce the watermark residual. In the detector part, a linear detection head is used for detecting the presence of watermarks, and a cross-attention module with the shared embedding table is used for message decoding. | 2502.04230v1 | 2 |
|
Figure 5:MUSHRA subjective listening test results comparing perceptual quality across different watermarking methods. Higher scores indicate better audio quality as rated by human listeners. Our method achieves quality scores comparable to AudioSeal. | 2502.04230v1 | 6 |
|
Figure 6:Validation accuracy and quality curve of different methods over training steps. Using a blended architecture,XAttnMarkis able to achieve a better balance in terms of learning efficiency in detection, and message-bit decoding and watermark imperceptibility. Compared to our method, WavMark, the fully-shared architecture, suffers from robustness degradation in detection as the training progresses; the fully-disjoint architecture, AudioSeal, suffers from training efficiency issues in learning message-bit decoding. | 2502.04230v1 | 7 |
|
Figure 7:Detection and Attribution Accuracy of our method on augmented samples across different augmentation strengths. | 2502.04230v1 | 8 |
|
Figure 8:The detection and attribution accuracy ofXAttnMarkacross different audio duration and watermark strengthα𝛼\alphaitalic_αon MusicCaps. The performance is averaged over all the standard audio editing. The attribution pool size is set to{100,1000,10000}100100010000\{100,1000,10000\}{ 100 , 1000 , 10000 }. | 2502.04230v1 | 9 |
|
Figure 9:Visualization of the original audio and watermark residuals across four different methods for a randomly selected sample from MusicCaps dataset, showing waveform (left), spectrogram (middle), and mel-spectrogram (right) representations. | 2502.04230v1 | 10 |
|
Figure 10:Visualization of the per-tile masking penalty distribution. The blue regions indicate tiles assigned a lower penalty for embedding a watermark. Our weighting scheme provides more fine-grained control. The audio sample is randomly selected from MusicCap(Agostinelli et al.,2023), and the watermarked audio is generated from a model at the very beginning of training. | 2502.04230v1 | 11 |
|
Figure 1.Illustration of the threat models. (a) Previous works assume that the data owner may be malicious and inject backdoors into the distilled dataset before distributing it to users. (b) In contrast, our threat model is more practical. We assume the data owner is benign. However, third parties, such as hackers or malicious users, may act maliciously. They could attack the system by hijacking the dataset distribution, injecting backdoors into the distilled dataset, and redistributing it to users. | 2502.04229v1 | 1 |
|
Figure 3.t-SNE visualization of the feature space. “Stars” and “Circles” represent the concept archetypes and real images, respectively. The reconstructed archetypes align closely with the deep feature representations of real images, effectively bridging the gap between the distilled data and real images. | 2502.04229v1 | 3 |
|
Figure 5.Visualization of benign and malicious distilled data. Without a direct comparison, users may struggle to sense the subtle differences due to the inherent abstraction of distilled datasets. | 2502.04229v1 | 11 |
|
Figure 1:A overall pipeline for SCP. During training, two augmented viewsTasuperscript𝑇𝑎T^{a}italic_T start_POSTSUPERSCRIPT italic_a end_POSTSUPERSCRIPTandTbsuperscript𝑇𝑏T^{b}italic_T start_POSTSUPERSCRIPT italic_b end_POSTSUPERSCRIPTof an image are generated from the dataset and processed by a frozen feature extractorf𝑓fitalic_fand a trainable cluster headg𝑔gitalic_g(a five-layer MLP). The objective is to minimize the cross-entropy loss between the outputs of the cluster headg𝑔gitalic_gfor the two augmented views. | 2502.04226v1 | 1 |
|
Figure 3:The visualization of clustering performance for SCP-CLIP withViT-B/32 backbone.(Left): An example of an image-to-image search on STL-10, showing clusters produced by CLIP (Top) and SCP (Bottom);(Right): Visualization of clustering performance. SCP-CLIP effectively enhances CLIP’s clustering performance through a cluster head. | 2502.04226v1 | 7 |
|
Figure 5:Comparison of different loss weightsα𝛼\alphaitalic_α, the solid line is the mean ARI across five runs, and the shaded region plots the standard deviation across our runs. | 2502.04226v1 | 13 |
|
Figure 7:SCP-CLIP Image-to-Image Search Pipeline depicted in Fig.3, utilizing the normalized outputs of the trained cluster headg(.)g(.)italic_g ( . )without applying softmax. | 2502.04226v1 | 15 |
|
Figure 8:The t-SNE representations learned by SCP-DINO on STL-10.ACC:98.4%:𝐴𝐶𝐶percent98.4ACC:98.4\%italic_A italic_C italic_C : 98.4 % | 2502.04226v1 | 16 |
|
Figure 9:The t-SNE representations learned by SCP-CLIP on STL-10.ACC:97.9%:𝐴𝐶𝐶percent97.9ACC:97.9\%italic_A italic_C italic_C : 97.9 % | 2502.04226v1 | 17 |
|
Figure 2:Meta architecture for ÉCLAIR showcasing the usage with two different (out of eight valid) prompts: Example a) uses the maximal information prompt to return bounding boxes along with their semantic class, markdown text, and tables and formulas. In b) we ask the model to return only markdown text without boxes or classes. All supported semantic classes are listed on the right. | 2502.04223v1 | 5 |
|
Figure S1:Illustrations of reading order over relevant text-like elements, i.e. Text, Section-header, List-item, Title and Formula. Other semantic classes (such as Picture, Footnote and Page-footer in the examples here) are not included in the reading order of the main body. (Note: We are not showing all the classes) | 2502.04223v1 | 7 |
|
Figure S3:Confusion matrix for ÉCLAIR boxes matched with ground truth on the DocLayNet evaluation dataset, averaged over thresholds ofIoU≥{0.5,0.55,…,0.9,0.95}𝐼𝑜𝑈0.50.55…0.90.95IoU\geq\{0.5,0.55,...,0.9,0.95\}italic_I italic_o italic_U ≥ { 0.5 , 0.55 , … , 0.9 , 0.95 }. | 2502.04223v1 | 11 |
|
Figure S4:PR-Curves for individual classes and averaged over all classes forIoU≥0.5𝐼𝑜𝑈0.5IoU\geq 0.5italic_I italic_o italic_U ≥ 0.5and 1001 recall-bins evaluated on the DocLayNet evaluation dataset. For ÉCLAIR, the scores are taken from the class-token logits. The mean precision and recall are taken fromTab.S1. | 2502.04223v1 | 12 |
|
Figure S5:Examples of pages with tables, formulae and pictures. On the left, predicted bounding boxes superimposed on the original sample image. On the right, the corresponding full predictions. | 2502.04223v1 | 13 |
|
Figure 1.An example of a predefined hierarchy used to raise the abstraction level of low-level event logs, where ”root” is the most abstract event andAssemblyLoader/Start_System_Threadingis the most specific, detailed event. | 2502.04219v1 | 1 |
|
Figure 1:The upper panel shows the general image stitching result and it can be successfully stitched. The lower panel shows the failed endoscopic image stitching result by using the traditional feature-based image stitching method. | 2502.04207v1 | 1 |
|
Figure 2:This figure illustrates an enhanced endoscopic image preprocessing pipeline designed for stitching endoscopic images. The process begins with the input endoscopic images, which undergo ellipse fitting to identify and mark regions of interest. Following ellipse fitting, the images are subjected to rotation adjustment to standardize their orientation. Specific unwrapping areas are then selected to facilitate detailed analysis. The unwrapping result provides a flattened view of the mucosal surface. Then, AHE is applied to the unwrapped images, enhancing features and improving the quality of the images for feature point matching. Finally, this process produces the stitching results. | 2502.04207v1 | 2 |
|
Figure 3:This figure presents a boxplot illustrating the number of valid matched feature points for keyframe selection across 20 videos using three different methods. The Original unwrapped images show the lowest number of valid matched feature points. In contrast, applying the AHE method to the unwrapped images markedly increases the number of feature points. The highest number of valid matched feature points is observed with the AHE Rotated unwrapped images, where both AHE and rotation methods are applied. The Wilcoxon signed-rank test is performed with AHE Rotated as the reference (“Ref”) method, to compare with other methods. “*” represents the significant (p<<<0.05) differences | 2502.04207v1 | 3 |
|
Figure 4:Result comparison among different image processing stages. The original unwrapped image shows only 5 valid matched feature points, resulting in a failed stitching outcome. The AHE unwrapped image, with 52 valid matched feature points, demonstrates improved stitching but still exhibits visible gaps between the images As shown by the yellow arrow. The AHE rotated unwrapped image, which has 57 valid matched feature points, achieves the best stitching result, with more seamless alignment as indicated by the yellow arrow. | 2502.04207v1 | 4 |
|
Figure 1:Packet delay distribution for LTT and QLTT in a resource allocation problem (adapted from[2]by using a different calibration data set𝒵𝒵\mathcal{Z}caligraphic_Z). | 2502.04206v1 | 1 |
|
Figure 1:Scatter plots of ASR to the ratio of the square root of the adversarial token suffix length in jailbreak attacks to the adversarial token suffix length during AT (i.e.,Mtest/Mtrainsubscript𝑀testsubscript𝑀train\sqrt{M_{\text{test}}}/M_{\text{train}}square-root start_ARG italic_M start_POSTSUBSCRIPT test end_POSTSUBSCRIPT end_ARG / italic_M start_POSTSUBSCRIPT train end_POSTSUBSCRIPT).
For each pair of base model and attack,48484848points are plotted.
A high ASR indicates a weak jailbreak robustness of the model. | 2502.04204v1 | 1 |
|
Figure 2:Curves of the ASR versus the adversarial suffix token length during AT (i.e.,Mtrainsubscript𝑀trainM_{\text{train}}italic_M start_POSTSUBSCRIPT train end_POSTSUBSCRIPT) under jailbreak attacks with different adversarial suffix token lengths (i.e.,Mtestsubscript𝑀testM_{\text{test}}italic_M start_POSTSUBSCRIPT test end_POSTSUBSCRIPT).Mtrain=0subscript𝑀train0M_{\text{train}}=0italic_M start_POSTSUBSCRIPT train end_POSTSUBSCRIPT = 0means that AT is not performed on the evaluated model.
A high ASR indicates a weak jailbreak robustness of the model. | 2502.04204v1 | 7 |
|
Figure 3:Utility analysis based on LC-WinRate.
A high LC-WinRate indicates strong model utility.
An LC-WinRate of50%percent5050\%50 %means that the evaluated model has the same quality as the reference model Davinci003. | 2502.04204v1 | 8 |
|
Figure 1:This figure shows that we expanded our training set from 435 to 7050 images for training EoE deep learning models. The left panel shows that the traditional AI model only uses rare pediatric EoE data. The right panel shows that we employ a wider range of data from different sources. | 2502.04199v1 | 1 |
|
Figure 2:This figure provides an overview of the learning framework. (a) illustrates the image classes, which include six EoE classes and five non-EoE classes, along with their approximate locations in the upper GI tract. (b) depicts the proportion of each class within the entire upper GI dataset. (c) displays the distribution of the EoE dataset after mining or collection. Finally, (d) shows how the DeiT AI model classifies these images. | 2502.04199v1 | 2 |
|
Figure 3:This figure shows the endoscopy EoE image and corresponding attention map. Image a shows an endoscopic image with label “exudate”, whose attention map is presented in image b. Image c presents an endoscopic image with “rings”, whose attention map visualization | 2502.04199v1 | 3 |
|
Figure 1:An overview of GRAPE. GRAPE takes multiple off-the-shelf existing datasets and optionally generates new responses, finds overlapping instructions with multiple different responses and selects responses that align with the base model’s distribution. This dataset is then used for standard supervised fine-tuning. | 2502.04194v1 | 1 |
|
Figure 2:The training loss curve of Llama3.1-8B onTulu-3-Olmo-2combo.Beststands for training with highest-probability responses (GRAPE-selected);Randomstands for randomly chosen responses for each instruction in the pool,Worststands for lowest probability responses in the base model’s distribution. Throughout training,Besthas lower loss thanRandomthanWorst. | 2502.04194v1 | 2 |
|
Figure 3:The average performance curve of Llama3.1-8B onTulu-3-Olmo-2combo.Beststands for training with highest-likelihood responses (GRAPE-selected);Randomstands for randomly chosen responses for each instruction in the pool. The average is taken over all items on table2except AlpacaEval2. | 2502.04194v1 | 3 |
|
Figure 1:The two major research questions we explore: (i) the grounding ability of pixel-level MLLMs in challenging scenarios (first row), (ii) the ability of vanilla MLLMs to perform grounding and when does it emerge (second row). Second row shows the noun phrase and its corresponding segmentation mask, highlighted in red, extracted from Llava 1.5(Liu et al.,2024)attention maps with three possible output masks to accommodate ambiguity in the point prompt, highlighted as a black dot. | 2502.04192v1 | 1 |
|
Figure 2:First shortcoming of pixel-level MLLMs is the degraded performance in visual question answering. The Predicted segmentation masks corresponding to the [SEG] token/s are highlighted in red. | 2502.04192v1 | 2 |
|
Figure 3:Second shortcoming of pixel-level MLLMs is the degraded performance in pixel-level visual grounding in certain models. The predicted segmentation is highlighted in red. | 2502.04192v1 | 3 |
|
Figure 4:Third shortcoming of pixel-level MLLMs is the degraded performance in instruction following, where the question is instructing the model to generate one letter from the options. | 2502.04192v1 | 4 |
|
Figure 9:The automatic selection baseline, PixFoundation, which uses a simple mechanism of highlighting the predicted masks in red then prompting a multi-modal large language model to select the right mask from the group of highlighted images, followed by the final mask selection. | 2502.04192v1 | 25 |
|
Figure 11:Normalized attention maps visualization showing the noun phrase and its corresponding attention in the output text for two PixMMVP examples using Cambrian-1 (8B) base MLLM. While the attention maps can not be directly used as segmentation, yet it provides initial locations for the maximally attended pixels corresponding to what the model is looking at. In certain scenarios it exactly aligns with the noun phrase describing it as in the two examples. Yet in certain scenarios as we showed earlier, the grounding of the referred expression in question emerges with other noun phrases describing it. | 2502.04192v1 | 39 |
|
Figure 12:First example of when grounding emerges, corresponding to Image 1 in Fig.17. Each row has the corresponding noun phrase on top and three potential SAM predicted masks highlighted in red using the maximum attention point of this noun phrase as a point prompt, highlighted as a black dot. It shows the output from mining the attention maps for pixel-level grounding using Llava 1.5 (7B) base MLLM. | 2502.04192v1 | 46 |
|
Figure 13:Third example of when grounding emerges, corresponding to Image 6 in Fig.17. Each row has the corresponding noun phrase on top and three potential SAM predicted masks highlighted in red using the maximum attention point of this noun phrase as a point prompt, highlighted as a black dot. It shows the output from mining the attention maps for pixel-level grounding using Llava 1.5 (7B) base MLLM. | 2502.04192v1 | 56 |
|
Figure 14:Second example of when grounding emerges, corresponding to Image 3 in Fig.17. Each row has the corresponding noun phrase on top and three potential SAM predicted masks highlighted in red using the maximum attention point of this noun phrase as a point prompt, highlighted as a black dot. It shows the output from mining the attention maps for pixel-level grounding using Llava 1.5 (7B) base MLLM. | 2502.04192v1 | 69 |
|
Figure 15:Fourth example of when grounding emerges, corresponding to Image 161 in Fig.17. Each row has the corresponding noun phrase on top and three potential SAM predicted masks highlighted in red using the maximum attention point of this noun phrase as a point prompt, highlighted as a black dot. It shows the output from mining the attention maps for pixel-level grounding using Llava 1.5 (7B) base MLLM. | 2502.04192v1 | 85 |
|
Figure 16:Fifth example of when grounding emerges, corresponding to Image 292 in Fig.17. Although the ground-truth segmentation is all background, since theoracleselection sees all the candidate masks with IoU zero, it defaults to the last selection. Each row has the corresponding noun phrase on top and three potential SAM predicted masks highlighted in red using the maximum attention point of this noun phrase as a point prompt, highlighted as a black dot. It shows the output from mining the attention maps for pixel-level grounding using Llava 1.5 (7B) base MLLM. | 2502.04192v1 | 98 |
|
Figure 17:Examples of the noun phrases and concept categories where the grounding emerged corresponding to the provided referring expression as input to the MLLM following the third evaluation protocol. The base MLLM used in these experiments are Llava 1.5 7B. | 2502.04192v1 | 114 |
|
Figure 1.Illustration of the MRAMG task (above), with scenarios below showing how integrating text and images enhances clarity and understanding. | 2502.04176v1 | 1 |
|
Figure 2.The construction pipeline of MRAMG-Bench. It is divided into three stages: (1) Data Selection and Preprocessing, we collect, clean and preprocess data to ensure the quality of initial datasets. (2) QA Generation and Refinement, we formulate questions and answers and refine the formulated QA pairs. (3) Data Quality Check, we engage annotators and experts in a three-stage data quality check to ensure high-quality of the entire benchmark. | 2502.04176v1 | 2 |
|
Figure 2:Accuracy scores for LS task performance. As token replacements increase, task performance decreases, but at different rates. Baseline performance shown in green. | 2502.04173v1 | 2 |
|
Fig. 2:Comparison of SMO and FNNLS[13]. In all instances of K (number of archetypes), the SMO updates converges to the optimal solution withinK2superscript𝐾2K^{2}italic_K start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPTiterations. The optimal solution for the differentK𝐾Kitalic_K’s are marked by the dotted black lines identified by FNNLS. | 2502.04172v1 | 8 |
|
Figure 1:Consumption patterns of different entities can be similar (entity 1 and entity 2). Single-task learning techniques can only leverage information from the corresponding entity, while multi-task learning techniques use information from multiple entities. | 2502.04163v1 | 1 |
|
Figure 4.Mexico City, daily, 6:00–9:00–12:00: time-elapsed effective distances of OD pairs. Top 0.95 quantile per day. Legend shows the frequency of occurrence in days of the OD pairs. Circled: on 08-22-2019, OD pair (9g3qy,9g3qv) with effective distance 10.2. | 2502.04162v1 | 8 |
|
Figure 6.Mexico City, daily, 6:00–9:00–12:00: time-elapsed effective distances of globally-unconnected OD pairs. Top 0.95 quantile per day. Legend shows the frequency of occurrence in days of such OD pairs. Circled: on 09-23-2019, OD pair (9g3w6,9g3w0) with effective distance 2.16. | 2502.04162v1 | 13 |
|
Figure 1:Annotation Highlights.We capture multi-day recordings of unscripted activities.Centre-Top: Recipes are recorded with steps and their preparation temporally annotated, along with ingredient addition. Ingredients are weighed and nutrition recorded.Centre-Middle: Dense fine-grained narrations detailing what, how, and why are parsed and clustered. Audio events are also annotated.Centre-Bottom: Object movements are temporally annotated with bounding boxes and hands and object masks.Right-Top: All annotations are temporally grounded in a 3D digital twin. We show trajectories of 3 (masked) objects:Sweet potato,Food processorandSpoon, highlighting relevant kitchen fixtures.Right-Bottom: Gaze captures when objects are primed (i.e.looked at) before being taken/placed. | 2502.04144v1 | 1 |
|
Figure 5:Frequency of verb clusters (top) and noun clusters (bottom) in narrated sentences by category, shown on a logarithmic scale. | 2502.04144v1 | 9 |
|
Figure 6:Nutrition is monitored throughout recipes as ingredients are incorporated into dishes. Here we show Banana Bread Chocolate Chip Cookies. We annotate when ingredients are weighed, document their nutrition, and locate their adding time, giving us overall dish nutrition at each stage. | 2502.04144v1 | 10 |
|
Figure 7:Digital Twin: from point cloud (left), to surfaces (middle) and labelled fixtures (right). We show two moved objects (masks on top) at fixtures:cheeseandpan.
Body poses from[85]. | 2502.04144v1 | 11 |
|
Figure 10:VQA Question Prototypes. We show our 30 question prototypes by category alongside the number of questions. Outer bars indicate the distribution over input lengths for each question. | 2502.04144v1 | 14 |
|
Figure 12:Effect of Input Length. Models struggle with questions of all video input lengths. s=second, m=minute,
h=hour. | 2502.04144v1 | 16 |
|
Figure 14:Long-Term VOS. jaccard index𝒥𝒥\mathcal{J}caligraphic_J& contour accuracyℱℱ\mathcal{F}caligraphic_Fshow Cutie and SAM2 struggle with segmenting objects. | 2502.04144v1 | 18 |
|
Figure A3:Per ingredient, we temporally label weighing (if present) and adding to recipe. Weighing is OCRed and manually verified. | 2502.04144v1 | 21 |
|
Figure A9:Left: Common source→→\rightarrow→destination object movements. Right: Common take/put locations of the top-20 moving objects. | 2502.04144v1 | 27 |
|
Figure A10:Object associations interface. For each query track, a set of possible object names extracted from the narrations is suggested, we provide these names as an additional annotations. Importantly, all previous tracks are sorted by their distance. It is trivial for the annotator to confirm that the track on the left, matches the box of chicken on the right, which has been positioned in exactly the same location (0.0m). By clicking ‘Add’, this track is corerctly associated connecting the object movement over time. | 2502.04144v1 | 28 |
|
Figure A11:Prediction Bias of Models. Most models have a bias in answer, although it is different for each model. | 2502.04144v1 | 29 |
|
Figure 1:Schematic of the two-microphone method above a baffled porous layer, flush-mounted with a rigid backing. | 2502.04143v1 | 1 |
|
Figure 2:Experimental setup of the two-microphone method in the anechoic chamber at the Marcus Wallenberg Laboratory for Sound and Vibration in Stockholm. | 2502.04143v1 | 2 |
|
Figure 5:Histograms of the recorded error per sample for the two-microphone method and for the neural network predictions using the numerical test set. | 2502.04143v1 | 5 |
|
Figure 3:Plot of one scenario of an infection from the diffusion SI-Equation. The Infection wave propagates over the nodes. | 2502.04140v1 | 8 |
|
Figure 4:Plot of some advection-field on the left side and an exemplary plot of the solution of the advection-diffusion equation over 4 time-steps. | 2502.04140v1 | 9 |
|
Figure 6:The dataset generated from the SI-diffusion equation over time. There are 400 lines plotted, each representing the values of an individual node over time. | 2502.04140v1 | 11 |
|
Figure 7:The SI-diffusion dataset(top) compared against the real-world German Influenza dataset(bottom). The step patterns in the Influenza dataset are presumably artefacts of the data collection/management pipeline. | 2502.04140v1 | 12 |
|
Figure 9:Illustration of the effect of additive noise sampled from the distribution𝒩(0,0.01)𝒩00.01\mathcal{N}(0,0.01)caligraphic_N ( 0 , 0.01 )from task 2 and task 4. On the left is the time series from a random graph node; on the right is its respective noise-added time series. The plotted section is the first1000100010001000time-steps of the training dataset. | 2502.04140v1 | 16 |
|
Figure 10:The RMSE over all samples from the test dataset. The less saturated colors represent the standard deviation for different initializations. | 2502.04140v1 | 17 |
|
Figure 13:Plots of the RMSEs of the presented models during the transfer-learning experiments. The arrows point from non-pre-trained to pre-trained models. | 2502.04140v1 | 24 |
|
Figure 1:(Left)Comparison of Shannon entropy, Rényi entropy, and behavioral entropy (ours) and their effects on dataset generation, shown in PHATE plots, when used as an exploration objective.(Right)Performance comparison of an offline RL algorithm (CQL) for three downstream tasks on datasets generated using Shannon, behavioral entropy (ours), and Rényi entropy for the parameterq=1.1𝑞1.1q=1.1italic_q = 1.1shown in the left-hand figure. | 2502.04141v1 | 1 |
|
(a)Perception of normal distribution (𝒩(0,1)𝒩01\mathcal{N}(0,1)caligraphic_N ( 0 , 1 )) under Prelec weighting function.Black dotted line shows standard normal density,f𝑓fitalic_f. Forα≈0𝛼0\alpha\approx 0italic_α ≈ 0the perceived density is uniform, indicatingover-weightingof uncertainty over entire support off𝑓fitalic_f. Forα≫0much-greater-than𝛼0\alpha\gg 0italic_α ≫ 0the perceived density approaches a step function around the mean, indicatingunder-weightingof uncertainty nearf𝑓fitalic_f’s tails.α=1𝛼1\alpha=1italic_α = 1recovers originalf𝑓fitalic_f. | 2502.04141v1 | 2 |
|
(a)Perception of normal distribution (𝒩(0,1)𝒩01\mathcal{N}(0,1)caligraphic_N ( 0 , 1 )) under Prelec weighting function.Black dotted line shows standard normal density,f𝑓fitalic_f. Forα≈0𝛼0\alpha\approx 0italic_α ≈ 0the perceived density is uniform, indicatingover-weightingof uncertainty over entire support off𝑓fitalic_f. Forα≫0much-greater-than𝛼0\alpha\gg 0italic_α ≫ 0the perceived density approaches a step function around the mean, indicatingunder-weightingof uncertainty nearf𝑓fitalic_f’s tails.α=1𝛼1\alpha=1italic_α = 1recovers originalf𝑓fitalic_f. | 2502.04141v1 | 3 |
|
(b)Diversity of Shannon, Rényi and BE values as a function of Bernoulli trial parameterp𝑝pitalic_p.Behavioral entropyHBsuperscript𝐻𝐵H^{B}italic_H start_POSTSUPERSCRIPT italic_B end_POSTSUPERSCRIPTcaptures entire behavior spectrum from overvaluing uncertainty (light blue,α≈0𝛼0\alpha\approx 0italic_α ≈ 0) to highly undervaluing uncertainty (dark blue,α≫0much-greater-than𝛼0\alpha\gg 0italic_α ≫ 0). Rényi,HRsuperscript𝐻𝑅H^{R}italic_H start_POSTSUPERSCRIPT italic_R end_POSTSUPERSCRIPT, captures the former (light red,q≈0𝑞0q\approx 0italic_q ≈ 0), but cannot capture the latter. Dotted red curve showsHRsuperscript𝐻𝑅H^{R}italic_H start_POSTSUPERSCRIPT italic_R end_POSTSUPERSCRIPTasq→∞→𝑞q\rightarrow\inftyitalic_q → ∞. Shannon,HSsuperscript𝐻SH^{\textup{S}}italic_H start_POSTSUPERSCRIPT S end_POSTSUPERSCRIPT, is dotted black curve. | 2502.04141v1 | 4 |
|
(b)PHATE plots for Renyi for Walker. Coverage forq=3.0,5.0𝑞3.05.0q=3.0,5.0italic_q = 3.0 , 5.0similar toq=2.0𝑞2.0q=2.0italic_q = 2.0. | 2502.04141v1 | 7 |
|
Figure 4:Comparison of offline RL performance over the entropy objectives used in dataset generation. Plots show mean and standard deviation over five seeds. Dotted line shows performance of RL policy trained online until approximate optimality. | 2502.04141v1 | 8 |
|
Figure 5:Offline RL results for allα𝛼\alphaitalic_αandq𝑞qitalic_qvalues evaluated. Initial trials showedq∈{2.0,3.0,5.0}𝑞2.03.05.0q\in\{2.0,3.0,5.0\}italic_q ∈ { 2.0 , 3.0 , 5.0 }led to performance no better (and usually worse) thanq=1.1𝑞1.1q=1.1italic_q = 1.1, so offline RL training for theseq𝑞qitalic_qvalues was not performed. | 2502.04141v1 | 14 |
|
Figure 6:Ablation result comparing the effect of performing 100K vs. 200K offline RL training steps on a 3M-element dataset generated using Shannon entropy as exploration objective. These results suggest that performing additional offline RL training has only a marginal effect on downstream task performance. | 2502.04141v1 | 20 |
|
Figure 13:Visualization of the BE reward function equation26by varying the parameterα𝛼\alphaitalic_αwithβ𝛽\betaitalic_βconditioned according to (4) from(Suresh et al.,2024)withM=512𝑀512M=512italic_M = 512, denoting the representation dimensions. These visualizations highlight the diversity and variety of rewards that can be obtained by a BE-maximizing reward function (blue region) as compared to the single SE objective (dotted black line). | 2502.04141v1 | 49 |
|
Figure 2:(a) The comparison of different query initialization methods.The FPS-based methods conduct farthest point sampling separately for each scene, placing more emphasis on positional information but lacking in aggregating content information. The learnable-based methods initialize a fixed number of queries for aggregating content information across all scenes, which is prone to empty sampling, thereby compromising foreground coverage. Our method leverages the advantages of both approaches to achieve a balanced and comprehensive solution.(b) The recall difference.The recall of the baseline shows instability during the iterative optimization process across layers, whereas our method, with the assistance of the Hierarchical Query Fusion Decoder, demonstrates a steady improvement in recall across each layer. | 2502.04139v1 | 2 |
|
Figure 3:The overall framework of our method BFL.The Agent-Interpolation Initialization Module is meticulously crafted to synergize the strengths of FPS and learnable queries, producing object queries better suited for complex and dynamic environments. The Hierarchical Query Fusion Decoder is utilized to retain low overlap queries that aid in recall rate. | 2502.04139v1 | 3 |
|
Figure 6:Visualization of instance segmentation results on ScanNetV2 validation set.The red boxes highlight the key regions. | 2502.04139v1 | 5 |
|
Figure 12:Parameter and runtime analysis of different methods on ScanNetV2 validation set.The runtime is measured on the
same device. | 2502.04139v1 | 8 |
|
Figure 14:The architecture of the transformer decoder layer.The figure is taken from the main text of Maft. | 2502.04139v1 | 9 |
|
Figure 15:Additional Visual Comparison on ScanNetV2 validation set.The red boxes highlight the key regions. | 2502.04139v1 | 10 |
|
Figure 16:Visual comparisons between the baseline and our method across different decoder layers on ScanNetV2 validation set.The red boxes highlight the key regions. | 2502.04139v1 | 11 |
|
Figure 17:Visual comparisons between the baseline and our method across different decoder layers on ScanNetV2 validation set.The red boxes highlight the key regions. | 2502.04139v1 | 12 |
|
Figure 18:Visual comparisons between the baseline and our method across different decoder layers on ScanNetV2 validation set.The red boxes highlight the key regions. | 2502.04139v1 | 13 |
|
Figure 1:HowGPT-4o miniresponds to the same question when the order of choices is reversed. The calls were made on Wednesday, January 29th, at 11:54 a.m. EST. | 2502.04134v1 | 1 |
|
Figure 1:Geometric intuition behind the mechanism of SIM with data from the toy model. Panela)depicts a binary classification problem which illustrates how training data can be oriented along manifolds of the formΦ=g(a,b)=abΦ𝑔𝑎𝑏𝑎𝑏\Phi=g(a,b)=abroman_Φ = italic_g ( italic_a , italic_b ) = italic_a italic_b. Panelb)shows the representation of the same data after applying SIM. The decision boundary between the two classes becomes simpler and, in this special case, the data evenbecomelinearly separable inΦΦ\Phiroman_Φ-space. | 2502.04131v1 | 1 |
|
Figure 4:Binary classification task for time series from the batch reactor model.
Displayed are 10 time series per class.
Observational noise simulated is normally distributed with standard deviationσ=3𝜎3\sigma=3italic_σ = 3. | 2502.04131v1 | 4 |
|
Figure 5:Experiment 1 showing improved classification with partially observed batch reactor model due to SIM.
Displayed are learning curves obtained from classifier training based on the fully observed (FO) dynamical model (dotted green) and the partially observed (PO) dynamical model, with and without application of SIM (marked with solid blue and dashed orange curves, respectively).
The training and test data used were generated on the dense time gridtdensesubscript𝑡denset_{\text{dense}}italic_t start_POSTSUBSCRIPT dense end_POSTSUBSCRIPTwith fixed observational noiseσ=1𝜎1\sigma=1italic_σ = 1on each observed component. | 2502.04131v1 | 5 |
|
Figure 6:Experiment 2 for the partially observed batch reactor model showing that SIM is robust to observational noise.
Displayed are generalization error and relative number of support vectors, each as a function of the observational noise.
Classification performance is compared for the partially observed (PO) dynamical model, with and without application of SIM (marked with solid blue and dashed orange curves, respectively).
Training and test data are generated on the dense time gridtdensesubscript𝑡denset_{\text{dense}}italic_t start_POSTSUBSCRIPT dense end_POSTSUBSCRIPTwithNtrain=20subscript𝑁train20N_{\text{train}}=20italic_N start_POSTSUBSCRIPT train end_POSTSUBSCRIPT = 20andNtest=400subscript𝑁test400N_{\text{test}}=400italic_N start_POSTSUBSCRIPT test end_POSTSUBSCRIPT = 400. | 2502.04131v1 | 6 |
|
Figure 7:Experiment 3 for the partially observed batch reactor model showing SIM is robust to changes in regularity and sparsity of the observed time series data.
Displayed are the learning curves obtained from classifier training for the partially observed (PO) dynamical model, with and without application of SIM (marked with solid blue and dashed orange curves, respectively).
The training and test data used were generated on the three different time gridstdense,tsparsesubscript𝑡densesubscript𝑡sparset_{\text{dense}},t_{\text{sparse}}italic_t start_POSTSUBSCRIPT dense end_POSTSUBSCRIPT , italic_t start_POSTSUBSCRIPT sparse end_POSTSUBSCRIPTandtirrsubscript𝑡irrt_{\text{irr}}italic_t start_POSTSUBSCRIPT irr end_POSTSUBSCRIPT, displayed in the left, middle and right panel, respectively.
Observational noise is fixed atσ=1𝜎1\sigma=1italic_σ = 1on each observed component. | 2502.04131v1 | 7 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.