image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
Figure 1:Training NNs as a first-passage process. Panel A presents the test accuracy as a function of the number of epochs for training without perturbations (orange line). Here,T𝑇Titalic_Tis the FPT to reach the target accuracy (dashed blue line). Panel B presents the test accuracy as a function of the number of epochs for training with a perturbation everyP𝑃Pitalic_Pepochs (orange line).TPsubscript𝑇𝑃T_{P}italic_T start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPTis the perturbed FPT to the target accuracy (dashed blue line) andτP(𝜽)subscript𝜏𝑃𝜽\tau_{P}(\boldsymbol{\theta})italic_τ start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT ( bold_italic_θ )is the residual time to reach the target accuracy after the first perturbation. | 2502.04121v1 | 1 |
|
Figure 2:An illustration of a QSS of a propagatorG(𝜽,t)𝐺𝜽𝑡G(\boldsymbol{\theta},t)italic_G ( bold_italic_θ , italic_t )at timest2>t1>trsubscript𝑡2subscript𝑡1subscript𝑡𝑟t_{2}>t_{1}>t_{r}italic_t start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT > italic_t start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT > italic_t start_POSTSUBSCRIPT italic_r end_POSTSUBSCRIPT. At QSS, the probability density functionϕ(𝜽)italic-ϕ𝜽\phi(\boldsymbol{\theta})italic_ϕ ( bold_italic_θ )(purple lines) maintains the same shape, while the fraction of models that survived (blue areas) declines over time through panels A-B, i.e.,ΨT(t1)>ΨT(t2)subscriptΨ𝑇subscript𝑡1subscriptΨ𝑇subscript𝑡2\Psi_{T}(t_{1})>\Psi_{T}(t_{2})roman_Ψ start_POSTSUBSCRIPT italic_T end_POSTSUBSCRIPT ( italic_t start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT ) > roman_Ψ start_POSTSUBSCRIPT italic_T end_POSTSUBSCRIPT ( italic_t start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT ). | 2502.04121v1 | 2 |
|
Figure 4:The mean trajectories of the test accuracy and their top and bottom deciles (shaded areas), that did not reach 72% test accuracy afterP𝑃Pitalic_Pepochs. From left to rightP=100𝑃100P=100italic_P = 100,50505050,20202020,10101010(see legend). The dashed black lines represent 72% test accuracy (the target). At epoch numberP𝑃Pitalic_Pwe apply the S&P protocol (upper row, panels A-E) or the partial resetting protocol (lower row, panels F-J). Panels E and J center the mean trajectories with respect to theirP𝑃Pitalic_P. The gray areas are from the moment of perturbation until a the first time that the average trajectory reaches the target accuracy. | 2502.04121v1 | 4 |
|
Figure 5:τ¯Psubscript¯𝜏𝑃\bar{\tau}_{P}over¯ start_ARG italic_τ end_ARG start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPTversus the perturbation time. The values ofτ¯Psubscript¯𝜏𝑃\bar{\tau}_{P}over¯ start_ARG italic_τ end_ARG start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPTfor S&P (orange circles) and partial SR (green triangles) are approximately constant whenP>τ¯P𝑃subscript¯𝜏𝑃P>\bar{\tau}_{P}italic_P > over¯ start_ARG italic_τ end_ARG start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT(outside the yellow area). Dashed dotted lines are averages of allτ¯P<Psubscript¯𝜏𝑃𝑃\bar{\tau}_{P}<Pover¯ start_ARG italic_τ end_ARG start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT < italic_P. | 2502.04121v1 | 5 |
|
Figure 6:The speedup gained by using different perturbation protocols versus the perturbation time. Orange and green solid lines are the theoretical predictions ofEq.9for S&P and partial SR, respectively. For the predictions, we used the value ofτ¯P∗subscript¯𝜏superscript𝑃\bar{\tau}_{P^{*}}over¯ start_ARG italic_τ end_ARG start_POSTSUBSCRIPT italic_P start_POSTSUPERSCRIPT ∗ end_POSTSUPERSCRIPT end_POSTSUBSCRIPTforP∗=100superscript𝑃100P^{*}=100italic_P start_POSTSUPERSCRIPT ∗ end_POSTSUPERSCRIPT = 100. We plot the predictions only down toP=τ¯P∗𝑃subscript¯𝜏superscript𝑃P=\bar{\tau}_{P^{*}}italic_P = over¯ start_ARG italic_τ end_ARG start_POSTSUBSCRIPT italic_P start_POSTSUPERSCRIPT ∗ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT.
The blue solid line is the theoretical prediction for the SR case (Eq.5). Symbols represent brute force training at every value ofP𝑃Pitalic_P. | 2502.04121v1 | 6 |
|
Figure 1.1:Evolution of GAN performance from 2014 to 2018 and 2024. The results for 2014–2018 are based on the demonstration by Goodfellow[1]at the International Conference on Learning Representations (ICLR) 2019 invited talk, showcasing the rapid advancements in GAN quality over the years[4,5,6,7]. The figure of 2024 from ISFB-GAN[8]. | 2502.04116v1 | 1 |
|
Figure 1.2:The basic architecture of a Generative Adversarial Network (GAN). The Generator creates fake images from random noise, while the Discriminator evaluates images to determine whether they are real or fake. Both networks are trained adversarially to improve the quality of the generated samples. | 2502.04116v1 | 2 |
|
Figure 5.1:Example images and their projected and re-synthesized counterparts. For each configuration, top row shows the target images
and bottom row shows the synthesis of the corresponding projected latent vector and noise inputs. With the baseline StyleGAN, projection
often finds a reasonably close match for generated images, but especially the backgrounds differ from the originals. Image from Karras et al.[15]in 2020 StyleGAN2 paper. | 2502.04116v1 | 5 |
|
Figure 10.1:The illustration of SeqGAN. Left:D𝐷Ditalic_Dis trained over the real data and the generated data byG𝐺Gitalic_G. Right:G𝐺Gitalic_Gis trained by policy gradient where the final reward signal is provided byD𝐷Ditalic_Dand is passed back to the intermediate action value via Monte Carlo search[188]. The figure from SeqGAN[185]. | 2502.04116v1 | 9 |
|
Figure 10.2:DCGAN uses small (5x5), twodimensional filters while WaveGAN uses longer (length-25), one-dimensional filters and a larger upsampling factor. Both strategies have the same number of parameters and numerical operations. The figure from WaveGAN[190] | 2502.04116v1 | 10 |
|
Figure 10.3:At each layer of the WaveGAN discriminator, the phase shuffle operation perturbs the phase of each feature map. The figure from WaveGAN[190] | 2502.04116v1 | 11 |
|
Figure 1:Adaptive margin from ambiguity. An illustration among (a) position embedding indicates per-point ambiguityaisubscript𝑎𝑖a_{i}italic_a start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPTcolored by a map ranging from00to1111, and (b) feature embedding yields similarities of intra-pairS+superscript𝑆S^{+}italic_S start_POSTSUPERSCRIPT + end_POSTSUPERSCRIPTand inter-pairS−superscript𝑆S^{-}italic_S start_POSTSUPERSCRIPT - end_POSTSUPERSCRIPT, using ambiguity-aware marginmisubscript𝑚𝑖m_{i}italic_m start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPTto adjust decision boundariesDB+𝐷superscript𝐵DB^{+}italic_D italic_B start_POSTSUPERSCRIPT + end_POSTSUPERSCRIPTandDB−𝐷superscript𝐵DB^{-}italic_D italic_B start_POSTSUPERSCRIPT - end_POSTSUPERSCRIPTin contrastive learning, which generates adaptive objectives to benefit embedding learning. | 2502.04111v1 | 1 |
|
Figure 3:Ambiguity visualization. A 3D point cloud scene is categorized by different semantic classes. We visualize the point-level ambiguity for each point, where the color from white to black indicates various ambiguity levels ranging in[0,1]01[0,1][ 0 , 1 ]. | 2502.04111v1 | 3 |
|
Figure 4:Visualization results on S3DIS (Area 5). The images from left to right are the input scene, ground truth of semantic labels, results predicted by PointNeXt, and our method. | 2502.04111v1 | 4 |
|
Figure 5:Visualization results on ScanNet. The images from left to right are the input scene, ground truth of semantic labels, results predicted by PointNeXt, and our method. | 2502.04111v1 | 5 |
|
Figure 2.Screenshot of the VTutor SDK demo athttps://codepen.io/AnonymousForReview/pen/xxvaxJg, the few lines of codes in the image are what developers need to do to add VTutor to their own website. | 2502.04103v1 | 2 |
|
Figure 1:(Left) Responses of the pretrained LLaVA to samples from TSI dataset (bottom) compared to DALL·E 2 generated images (top) for the‘cooking on a stove’class. (Right) LLaVA’s correct response to the same TSI image after fine-tuning LLaVA using LoRSU. | 2502.04098v1 | 1 |
|
Figure 3:TFlops and trainable parameters comparison between LoRSU with CLIP loss (LoRSU), perplexity loss (LoRSU-Ppl), and LoRA-F. | 2502.04098v1 | 3 |
|
Figure 1:Diagram of our protocol. The data from the PETBlscan performed before the start of treatment are used to quantify the initial metabolism of the tumor. The data from the PETFuscan performed after the first course of chemotherapy are used to assess residual metabolism of the tumor and the metabolic response compared to the results of PETBl. Magnetic Resonance Imaging (MRI) exam, which is an addition to the main protocol, was performed at baseline step. | 2502.04083v1 | 1 |
|
Figure 2:Axial PET images of breast tumorabefore andbafter first cycle of NAC for the same patient. The red arrow indicates the tumor lesion. | 2502.04083v1 | 2 |
|
Figure 3:Pipeline of the segmentation of the tumor on PETBlscans. The pre-processing part explains how the input data are organized and cleaned before being fed into the network for the segmentation of PETBlscans alongside the ground truth. | 2502.04083v1 | 3 |
|
Figure 4:Pipeline for the management of the PETFuscans and the biomarker extraction:aExtraction of biomarkers using the segmented mask at the baseline level.bSegmentation of the PETFuscans using the fine-tuned baseline model (modelBL).cCalculation of changes in the biomarkers between PETFuand PETBl, such as SUVmax, to observe the impact of NAC. Active learning process is a process in which outliers identified in the PETFusegmentation by the quality control system are then manually labeled to further refine the model. The terme ”Mapping” corresponds to the extraction of the biomarkers from the region of interest associated to a segmented mask. | 2502.04083v1 | 4 |
|
Figure 5:The distribution of MTV in PETBlscans, calculated from labels, is compared to the MTV ratio between PETFuand PETBlcasesabefore andbafter fine-tuning. The data points encircled by a red line are identified as outliers, while those within a rectangular green boundary are considered as extreme outliers, being the farthest from the threshold. The blue perpendicular dotted line to the x-axis represents the threshold. The regression line inaandbrepresents the relationship between the MTV of labeled scans and the MTV ratio between PETBland PETFuscans. | 2502.04083v1 | 5 |
|
Figure 6:Segmentation results for two patients at both baseline and follow-up stages. For the patient 1,aanddrepresent the PET images,bandedisplay the corresponding ground truths in red overlaid on the images (a) and (d) respectively.candfshow the automatically segmented masks in red for images (a) and (d) respectively. Moving on to the patient 2,gandjdepict the PET images,handkdisplay the corresponding ground truth in red overlaid on the images (g) and (j) respectively, andiandlshow the automatically segmented masks in red for images (g) and (j) respectively. | 2502.04083v1 | 6 |
|
Figure 7:Example of a quality control method from predicted tumor areas from PETBland PETFufrom the same patient. In theValidated scenario on the left, both ROIs automatically predicted are located in the same quadrant, with the MTV ratio between them being below the specified threshold. In theInvalidated scenario on the right, although the ROIs are in the same quadrant, the MTV ratio exceeds the threshold. This is one of the extreme outliers identified by our quality control system. | 2502.04083v1 | 7 |
|
Figure 8:Box plots comparing biomarker values for SUVmax, MTV, and TLG at baseline and follow-up stages for 180 patients. Each box represents the interquartile range (IQR), with the red line inside indicating the median value. Measurements at baseline stage (left) are compared to measurements at follow-up stage (right) for each parameter, illustrating changes after NAC based on the biomarker values extracted from our system. | 2502.04083v1 | 8 |
|
Figure 1:DEAL generates a sequence of reconstructions𝐱ksubscript𝐱𝑘{\bf{x}}_{k}bold_x start_POSTSUBSCRIPT italic_k end_POSTSUBSCRIPTvia (6) from the inputs𝐲𝐲{\bf{y}}bold_yand𝐇𝐇{\bf{H}}bold_H, initalization𝐱0=𝟎subscript𝐱00{\bf{x}}_{0}={\bf{0}}bold_x start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT = bold_0, and hyper-parametersσ𝜎\sigmaitalic_σandλ𝜆\lambdaitalic_λ.
When the stop condition is met, it returns𝐱^^𝐱\hat{{\bf{x}}}over^ start_ARG bold_x end_ARG. | 2502.04079v1 | 1 |
|
Figure 1:A comparison of H2O, Quest, and AttentionPredictor for identifying critical tokens in the next step with history attention score. Our learning-based spatiotemporal predictor captures the dynamic attention patterns and accurately predicts next-step attention scores. | 2502.04077v1 | 1 |
|
Figure 3:Overview of AttentionPredictor and cross-token prefetching framework. (a)AttentionPredictorformulates the attention history as a spatiotemporal sequence, and predicts the attention at the next step with a pre-trained model. To enhance efficiency, the attention history is updated in a compressed form at each decoding step. (b)The cross-token prefetching frameworkasynchronously evaluates critical tokens and fetches KV for the next token during the LLM inference, thereby accelerating the decoding stage. | 2502.04077v1 | 3 |
|
Figure 4:Timeline of our proposed cross-token prefetching. By asynchronously loading the critical KV cache for the next token, our framework hides the token evaluation and transfer latency, accelerating the decoding stage of LLM inference. | 2502.04077v1 | 4 |
|
Figure 1:Comparison of concurrent and previous AIGC videos. Videos are generated by Lavie(Wang et al.,2023c)(1st row) and Sora(Brooks et al.,2024)(2nd row), respectively. Nouns that should be present in the video are highlighted in orange, while adjectives with more details are highlighted in blue. The new-generation AIGC videos contain richer content. | 2502.04076v1 | 1 |
|
Figure 7:The ranking of next-generation models provided by models trained on different AIGC VQA datasets. | 2502.04076v1 | 7 |
|
Figure 8:Scatter plots of the predicted scores and ground-truth MOSs. A brighter scatter point represents higher density. | 2502.04076v1 | 8 |
|
Figure 1:When asking questions to a LLM, almost all models will answer the user’s question "politely" as shown in the figure, but when we apply our emotion vector, the model will produce strong emotional expressions. The example in the figure uses the llama3.1-8B-Instruct model and applies the extracted anger vector. More detailed examples are shown in Table1. | 2502.04075v1 | 1 |
|
Figure 15:A form is expected to behave dynamically by providing real-time feedback to users through form validation. | 2502.00226v1 | 15 |
|
Figure 16:A useEffect hook should be used to re-evaluate the form’s validity whenever any error or input changes. | 2502.00226v1 | 16 |
|
Figure 18:The handleChange function should update the error state dynamically based on user input to ensure the UI always reflects the correct validation status. | 2502.00226v1 | 18 |
|
Figure 21:The conversion logic should ensure that the corresponding input field is updated whenever a change occurs. | 2502.00226v1 | 21 |
|
Figure 22:The ngOnInit() lifecycle hook initializes the component’s state. Initially, the input1 and input2 should be set to null, with the default units being Kilometer and Meter. | 2502.00226v1 | 22 |
|
Figure 23:Fix: Adding the current_user method to the AuthenticateUser concern to ensure that the method is available for use in the controller actions. | 2502.00226v1 | 23 |
|
Figure 24:Fix: Adding validations ensured that only valid job entries were created, improving data integrity. | 2502.00226v1 | 24 |
|
Figure 25:Fix: The response was updated to include details about the associated job and user, making it more comprehensive. | 2502.00226v1 | 25 |
|
Figure 1:MAB exploit puzzle forGpt-4(left),Gpt-4with CoT (middle), andGpt-3.5with CoT (right), all with "buttons" prompt. The following conventions apply to all figures in this section. Each line corresponds to a particular value ofT𝑇Titalic_Tand plotsFracCorrect(ϵ,T)FracCorrectitalic-ϵ𝑇\mathrm{FracCorrect}(\epsilon,T)roman_FracCorrect ( italic_ϵ , italic_T )against empirical gapϵitalic-ϵ\epsilonitalic_ϵon the X-axis. The shaded band around the line represents a95%percent9595\%95 %confidence interval. The dashed line is the number of tasks (“runs") with empirical gap at mostϵitalic-ϵ\epsilonitalic_ϵ; the corresponding Y-scale is on the right. | 2502.00225v1 | 1 |
|
Figure 2:Gpt-4succeeds on a small CB exploit puzzle (left) and fails on a slightly larger one (right). | 2502.00225v1 | 2 |
|
Figure 3:CB exploit puzzle withd=K=2𝑑𝐾2d=K=2italic_d = italic_K = 2andT=4000𝑇4000T=4000italic_T = 4000:mitigations help substantially.Gpt-4without CoT (left) andGpt-4with CoT (right). Note that providing the full history with thisT𝑇Titalic_Tvastly exceeds the context window forGpt-4,Gpt-4o, andGpt-3.5. | 2502.00225v1 | 5 |
|
Figure 4:CB exploit puzzle withd=K=5𝑑𝐾5d=K=5italic_d = italic_K = 5andT=1000𝑇1000T=1000italic_T = 1000:mitigations perform badly, but (mostly) much better than the no-mitigation baseline.Gpt-4owithout CoT. | 2502.00225v1 | 8 |
|
Figure 5:Gpt-4oon the text-based CB exploit puzzle:
somemitigationshelp, but areoutperformed by linear regression. | 2502.00225v1 | 9 |
|
Figure 6:Algorithm’s average expected reward𝚛𝚎𝚠¯(task,K)¯𝚛𝚎𝚠task𝐾\overline{\mathtt{rew}}(\text{task},K)over¯ start_ARG typewriter_rew end_ARG ( task , italic_K )(averaged over rounds and over runs), againstK𝐾Kitalic_K, the number of candidates. Each line corresponds to a prompting strategy or the𝚁𝚊𝚗𝚍𝚘𝚖𝚁𝚊𝚗𝚍𝚘𝚖\mathtt{Random}typewriter_Randombaseline.
The shaded regions represent a95%percent9595\%95 %confidence interval. | 2502.00225v1 | 10 |
|
Figure 7:Arm histogram for one particular question (“What is the meaning of freedom?”, left) and averaged over all questions (right). We considerK=10𝐾10K=10italic_K = 10suggestions generated by our one-by-one prompt without encouragement. For a given “run" of the candidate selection, the suggestions are ranked by the expected reward, and then thei𝑖iitalic_i-th expected reward is averaged over all runs, for eachi∈[K]𝑖delimited-[]𝐾i\in[K]italic_i ∈ [ italic_K ]. | 2502.00225v1 | 13 |
|
Figure 8:Algorithm’s average expected reward𝚛𝚎𝚠¯(category,K)¯𝚛𝚎𝚠category𝐾\overline{\mathtt{rew}}(\text{category},K)over¯ start_ARG typewriter_rew end_ARG ( category , italic_K )(averaged over rounds and tasks), againstK𝐾Kitalic_K, the number of candidates. Each line corresponds to a prompting strategy or the Category-Only baseline.
The shaded regions represent a95%percent9595\%95 %confidence interval. A single arXiv category (“General Relativity and Quantum Cosmology", left), averages over 6 categories (right). | 2502.00225v1 | 16 |
|
Figure 10:Cumulative fraction correct forGpt-4with chain-of-thought reasoning in the MAB adverts puzzle. | 2502.00225v1 | 20 |
|
Figure 13:Results averaged over the first six questions, for embeddings generated using the universal sentence encoder. | 2502.00225v1 | 23 |
|
Figure 1:Scheme of a BCI stimuli-based-on system. The signal is recorded by
means of certain acquisition technique, for instance the EEG recording. Then
the digital signal obtained is processed (filtered
+ feature extraction + classification) in order to be parsed into commands.
Finally, this extracted information (the commands) will be send through an
interface with the different systems or devices to do the desired actions
(system A and/or system B in figure). These systems can be for instance a
commanded wheelchair, a BCI speller to send messages, a commanded robotic
arm, etc. | 2502.00220v1 | 1 |
|
Figure 2:Scheme of six P300-ERPs taken from the II BCI Competition problem 2b (speller matrix) dataset. Each interval, of 600ms, was taken after a stimulus was shown to the subject and the intensified row or column belong to the target character. The P300-ERP should manifest an increment of amplitude around 300ms after the stimulus. | 2502.00220v1 | 2 |
|
Figure 3:Scheme of a P300 BCI system. The subject pays attention to a speller matrix program in a screen, in which the different columns and rows intensifies randomly. At the same time, the brain activity is recorded through EEG. As a result of row or column intensification of the focused character (marked as * in figure4) the signal should increment its amplitude. This increment appears after 300ms and does not last longer that 600ms. | 2502.00220v1 | 3 |
|
Figure 4:Structure used in the recording session showed in figure3. This figure is a reduced but more detailed version of
figure2. Each numbered source correspond to each stimulus
(rows 1-6, and columns 7-12) of the speller matrix, where each pulse
corresponds to an intensification of that row or column. Once the infrequent
stimulus appears, an amplitude increment should appear in the recorded
signal 300 ms. The change generated by the ERP should not continue 600ms
after the stimulus. After the marked row and column should manifest a P300-ERP
in the figure. In this case, the subject was paying attention to the
character “E” according to figure3. | 2502.00220v1 | 4 |
|
Figure 6:Sample of a CompLearn’s hierarchical tree[54]from a subset of objects, of the second BCI Competition. CompLearn is a
compression based toolbox that takes, in this case, the NCD distance matrix as
input and generates a hierarchical tree. The configuration used wasC=M=4𝐶𝑀4C=M=4italic_C = italic_M = 4, according to the method described in figure5.
The blue nodes correspond to the P300s objects and the black ones to the
non-P300. The Silhouette Coefficient of the dendrogram is 0.3474. | 2502.00220v1 | 6 |
|
Figure 7:Sample of a PEx’s projection from a subset of objects, of the second BCI Competition. PEx is a
visualization tool that takes, in this case, the NCD distance matrix as
input and generates a projection through mapping techniques.[56,53].
Following the
method described in figure5, we take the same
object configuration of figure6, whereM=C=4𝑀𝐶4M=C=4italic_M = italic_C = 4. The
white nodes correspond to the P300s objects and the black ones to the
non-P300s. The Silhouette Coefficient of the projection is 0.1352 | 2502.00220v1 | 7 |
|
Figure 8:Output of PEx from different object configurations, for a single
electrode
(Cz according to[60]). We observed
in this experiment that the increment of segments per object seems to
improve the separation of both clusters. In order to compare each
configuration we take a subset of objects for each one of them due to the
last configuration, Lower-right, that can not generate more than 18 P300 objects. | 2502.00220v1 | 8 |
|
Figure 9:Simulation over the number of segments per object (single P300-ERPs
or non-P300-ERPs) for an unique electrode
(Cz according to[60])
through Maketree’s clustering. Each data point represents the average of
iterating 100 times the simulation (20 files per simulation due to cost
limitations of Maketree algorithm). The data was taken
from both training sessions (10 and 11). The iterated parameters were the
number of segments used in both concatenation and average processes. In this
case, the curve only represents those configurations with equal number of
segments for both concatenation and average processes (i.e.M=C𝑀𝐶M=Citalic_M = italic_C). | 2502.00220v1 | 9 |
|
Figure 10:Silhouette Coefficient distribution obtained for each electrode for
both sessions 10 and 11. Each score was calculated from aM=C=8𝑀𝐶8M=C=8italic_M = italic_C = 8configuration from 100 iterations of CompLearn simulation. The upper figure
shows the SC distribution among electrode through boxplots. The lower figure
shows the median of the SC for each electrode across the scalp. | 2502.00220v1 | 10 |
|
Figure 11:Silhouette Coefficient boxplot distribution obtained for each electrode for sessions 10 and 11, individually. Each score was calculated from aC=M=8𝐶𝑀8C=M=8italic_C = italic_M = 8configuration from 100 iterations of CompLearn simulation. | 2502.00220v1 | 11 |
|
Figure 12:Silhouette Coefficient means distribution across the scalp obtained for each electrode for sessions 10 (left) and 11 (right), individually. Each score was calculated from aC=M=8𝐶𝑀8C=M=8italic_C = italic_M = 8configuration from 100 iterations of CompLearn simulation. | 2502.00220v1 | 12 |
|
Figure 13:Grid simulation over the number of ERPs used in both the average and
concatenation processes. In the first color map, the first 8 electrodes with
higher SC were used to perform the simulation. These 8 electrodes, also,
correspond to the center-frontal area of the scalp. The other 6 color maps show
the results for the following 6 subsets of 8 electrodes, from higher SC to
lower. The black square shows the best configuration obtained in the entire simulation. | 2502.00220v1 | 13 |
|
Figure 14:Dendrogram (using CompLearn toolkit) obtained from a
subset of optimized objects. The blue nodes correspond to P300s objects and
the black ones to non-P300. The Silhouette Coefficient of the dendrogram is
0.56. Following the scheme of figure5, we take a
configuration for objects creation ofM=10𝑀10M=10italic_M = 10andC=12𝐶12C=12italic_C = 12. This configuration corresponds to the optimal separation of P300 and non-P300 in figure13. Here, we can see how the separation of P300 and non-P300 from figure6, where the number of means and concatenations was not optimized, is improved considerably. | 2502.00220v1 | 14 |
|
Figure 15:PEx projection from a subset of optimized objects. The white nodes
correspond to the P300 objects and the black ones to non-P300. The
Silhouette Coeficient of the projection is 0.36. Following the scheme of
figure5, we take a configuration ofM=10𝑀10M=10italic_M = 10andC=12𝐶12C=12italic_C = 12. This configuration
corresponds to the optimal separation of P300 and non-P300 in figure13. Here, we
can see how the separation of P300 and non-P300 from figure7, where the
number of means and concatenations was not optimized, is improved considerably. | 2502.00220v1 | 15 |
|
Figure 16:Silhouette Coefficient means distribution across the scalp obtained
for each electrode for the subject A and B, respectively, of the third BCI Competition. Each
score was calculated from aC=M=8𝐶𝑀8C=M=8italic_C = italic_M = 8configuration with CompLearn. | 2502.00220v1 | 16 |
|
Figure 17:PEx projection from Subject A and B, respectively, obtained from the top 3 central-frontal
subset of optimized objects from the BCI
Competition III dataset II. The white nodes
correspond to the P300 objects and the black ones to non-P300. The
Silhouette Coeficient of the projections are 0.25 and 0.38, respectively.
The configuration isM=10;C=12formulae-sequence𝑀10𝐶12M=10;C=12italic_M = 10 ; italic_C = 12. | 2502.00220v1 | 17 |
|
Figure 1:Toy Experiment.The four plots on the left-side visualize the loss trajectories of various MTL methods
from 5 initialization points (∙∙\bullet∙) on a toy 2-task learning problem (seeSection4.1andAppendixCfor more details).
Trajectories transition from blue to green, indicating progress over time.
All 5 initialization points for FAMO reach the Pareto front (gray curve),
while and 3 for NashMTL and all 5 for both CAGrad andConicGradreach the global minima (★★\bigstar★)
withConicGradconverging significantly faster.
The plot on the far-right compares the convergence speeds over training steps,
showing thatConicGradachieves the lowest loss (dashed black line) faster than all competing methods. | 2502.00217v1 | 1 |
|
Figure 3:Visualizing Conic vs. Directional Constraints.We visualizeConicGradand CAGrad(Liu et al.,2021a)constraints in a toy setup. The x and y axes denote all possible direction vectors in2222D spaceℝ2superscriptℝ2\mathbb{R}^{2}blackboard_R start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT, and the plot indicates which vectors in this space satisfyConicGradandCAGradconstraints. | 2502.00217v1 | 3 |
|
Figure 4:Scalability Experiments on CelebA.We measure the computational overhead of MTL methods as the model size increases (in terms of number of parameters) to illustrate how these methods scale. | 2502.00217v1 | 4 |
|
Figure 5:Contour Plots ofc𝑐citalic_candγ𝛾\gammaitalic_γon three MTL benchmarks.We ablate the hyperparametersγ∈[0.001,0.01]𝛾0.0010.01\gamma\in[0.001,0.01]italic_γ ∈ [ 0.001 , 0.01 ]on the x-axes
andc∈{0.1,0.25,0.5,0.75,0.9}𝑐0.10.250.50.750.9c\in\{0.1,0.25,0.5,0.75,0.9\}italic_c ∈ { 0.1 , 0.25 , 0.5 , 0.75 , 0.9 }on the y-axes.
The raw data consists of discrete values forγ𝛾\gammaitalic_γandc𝑐citalic_cat specific points,
and we use interpolation to fill in the gaps
to create a continuous surface that reveals howΔm%Δpercent𝑚\Delta m\%roman_Δ italic_m %(darker areas indicate better performance) varies across the hyperparameter space. | 2502.00217v1 | 5 |
|
Figure 6:Geometric Interpretation ofConicGrad.The black arrow indicates the main objective gradients vectorg0subscript𝑔0g_{0}italic_g start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT, and the region covered inpinkindicates the valid region. We plot several values ofc𝑐citalic_cto visualize how the region of admissible update direction vectors changes. | 2502.00217v1 | 6 |
|
Figure 7:Loss Landscape of the Toy Example.(a) First objective,ℒ1(θ)subscriptℒ1𝜃\mathcal{L}_{1}(\theta)caligraphic_L start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT ( italic_θ );
(b) Second objective,ℒ2(θ)subscriptℒ2𝜃\mathcal{L}_{2}(\theta)caligraphic_L start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT ( italic_θ );
(c) The overall objective,12(ℒ1(θ)+ℒ2(θ))12subscriptℒ1𝜃subscriptℒ2𝜃\frac{1}{2}\big{(}\mathcal{L}_{1}(\theta)+\mathcal{L}_{2}(\theta)\big{)}divide start_ARG 1 end_ARG start_ARG 2 end_ARG ( caligraphic_L start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT ( italic_θ ) + caligraphic_L start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT ( italic_θ ) ); and
(d) 3D plot of (c).
The global minima for (a), (b), and (c) are represented by★★\bigstar★,
and the Pareto front for (c) is denoted by a gray line.
Balancing the conflicting gradients while navigating the overall objective is crucial;
since without it, the optimization process is prone to getting stuck in either of the two suboptimal local minima. | 2502.00217v1 | 7 |
|
Figure 1:Correlation between the gradient norm and the maximum Hessian eigenvalue. Each point represents a parameter block (pre-trained RoBERTa on RTE). | 2502.00213v1 | 1 |
|
Figure 2:Correlation between the full-batch gradient and gradient error. Each point represents the absolute values of a coordinate (pre-trained RoBERTa on RTE). | 2502.00213v1 | 2 |
|
Figure 5:Linear head norm of fine-tuned ResNet18 on Flowers102, normalized by parameter dimension. Error bars indicate standard deviations. | 2502.00213v1 | 11 |
|
Figure 1:Self-play Theorem Prover (STP).
Our model simultaneously takes on two roles — the conjecturer that generates new, related conjecture given a seed theorem with proof (Step 1), and the prover that attempts to prove the statements in an existing dataset and the generated conjectures (Step 2). Step 4 selects the correct, approachable, elegant, yet challenging conjectures to train the conjecturer, and the verifier selects correct proofs in Step 3 to train the prover.
The main difference between STP and expert iteration is the conjecturer role highlighted with a yellow background. | 2502.00212v2 | 1 |
|
Figure 2:The cumulative pass rates of STP, expert iteration, and parallel sampling on LeanWorkbook shows that STP achieves a much better scaling in terms of the performance vs number of generated proofs. The compute for generating conjectures and training the conjecturer in STP is negligible because the number of generated proofs during training is 64 times the number of conjectures. | 2502.00212v2 | 2 |
|
Figure 2:The cumulative pass rates of STP, expert iteration, and parallel sampling on LeanWorkbook shows that STP achieves a much better scaling in terms of the performance vs number of generated proofs. The compute for generating conjectures and training the conjecturer in STP is negligible because the number of generated proofs during training is 64 times the number of conjectures. | 2502.00212v2 | 3 |
|
Figure 3:Comparison of pass rates on miniF2F-test (y-axis) with different numbers of inference-time samples (x-axis). The model trained with STP consistently outperforms the DeepSeek-Prover-V1.5 series. | 2502.00212v2 | 4 |
|
Figure 5:Left:Comparison of pass rates between STP, two implementations of expert iteration, and parallel sampling methods on LeanWorkbook.Right:Comparison of pass rates between STP and baseline methods on LeanWorkbook, starting with the SFT model. The red crosses shows the points where we refresh the self-play training as described in Section4.1. | 2502.00212v2 | 6 |
|
Figure 1:Illustration of the proposed lightweight, computationally- and energy-efficient EcoWeedNet model’s performance in sustainable (low-carbon) consumer electronics-based precision agriculture applications for automated inspection and detection of weeds. | 2502.00205v1 | 1 |
|
Figure 5:Comparative visualization of original images, baseline YOLOv11n, and EcoWeedNet feature responses. | 2502.00205v1 | 9 |
|
Figure 6:Confusion matrix from our enhanced model showcasing the classification accuracy and distribution of results. | 2502.00205v1 | 10 |
|
(a)Cumulative utility ofAlgorithm1instantiated with OFUL (Algorithm1-OFUL), Algorithm3333ofHarris et al. [14], and the random baseline overT=2,000𝑇2000T=2,000italic_T = 2 , 000rounds in a setting with5555follower types, where each player has3333actions and the context dimension is also3333. Results are averaged over10101010runs. The hyperparameter of Algorithm3333ofHarris et al. [14]was tuned to maximize performance. | 2502.00204v1 | 1 |
|
(a)Cumulative utility ofAlgorithm1instantiated with OFUL (Algorithm1-OFUL), Algorithm3333ofHarris et al. [14], and the random baseline overT=2,000𝑇2000T=2,000italic_T = 2 , 000rounds in a setting with5555follower types, where each player has3333actions and the context dimension is also3333. Results are averaged over10101010runs. The hyperparameter of Algorithm3333ofHarris et al. [14]was tuned to maximize performance. | 2502.00204v1 | 2 |
|
(b)Cumulative utility ofAlgorithm1instantiated with OFUL (Algorithm1-OFUL) and the random baseline overT=2,000𝑇2000T=2,000italic_T = 2 , 000rounds in a setting with4444follower types, where each player has4444actions and the context dimension is also4444. Results are averaged over10101010runs. Algorithm3333ofHarris et al. [14]is not applicable in this setting because the follower’s utility depends on the context. | 2502.00204v1 | 3 |
|
Figure 1:The average reward (left) and win-rate (mid) overlmsys (valid)prompts along training. Therightfigure shows the MT bench (judged by Mistral Large 2). Error bars represent 95% confidence intervals over 3 independent runs. We compare two training datasets, which are generated by the llama3-8b-sft model usinglmsysandsyntheticprompts, respectively. We observe training on in-distributionlmsysprompts achieves higher rewards than training on out-of-distributionsyntheticprompts. However, the MT-Bench metric has a large variance, hardly showing any learnings. | 2502.00203v1 | 1 |
|
Figure 2:Online RPO-bwdvsonline RPO-sqloo(RLOO). We plot average rewards onlmsys(valid)(left) and the KL divergence with the reference policy (right). The valid reward increases faster and the KL divergence increases slower forRPO-bwd. This indicates thatonline RPO-bwdcan better optimize the RLHF objective (Eq1) thanRLOO. In addition,RLOO’s training exploded in the middle; whileRPO-bwd’s training kept stable in all our runs. | 2502.00203v1 | 2 |
|
Figure 7:Yearly Trends of Machine Learning Algorithm Application in Fraud Detection from 2019 to 2024. | 2502.00201v1 | 7 |
|
Figure 1:Hypotheses that are a priori preferred by the learnerandhave low risk form a set offeasible hypotheses. Feasible hypotheses are of major interest because hypotheses inℋEx∖ℋFesuperscriptℋExsuperscriptℋFe\mathcal{H}^{\mathrm{Ex}}\setminus\mathcal{H}^{\mathrm{Fe}}caligraphic_H start_POSTSUPERSCRIPT roman_Ex end_POSTSUPERSCRIPT ∖ caligraphic_H start_POSTSUPERSCRIPT roman_Fe end_POSTSUPERSCRIPTcould be easily eliminated during learning. | 2502.00197v1 | 1 |
|
Figure A1:We use a holistic term — learning under distributional shift (LInvsuperscript𝐿InvL^{\text{Inv}}italic_L start_POSTSUPERSCRIPT Inv end_POSTSUPERSCRIPT) — to capture the focus on invariance and the static nature of the optimal hypothesis.a.In-domain PAC-learning is the most basic type of learning.b-g.Sub-frameworks encompassed by “learning under distributional shift”.h.A compact and unified diagram for “learning under distributional shift”. | 2502.00197v1 | 3 |
|
Figure A2:Schematic illustration of streaming, continual and lifelong learning, all featuring a progressive manner of receiving data and inferring optimal hypotheses.a.New data arrive in individual examples and in batches for streaming and continual learning, respectively, which is a minor aspect that we do not distinguish in the diagrams.b.Lifelong learning extends continual learning by additionally requiring an explicit expansion ofℋExsuperscriptℋEx\mathcal{H}^{\mathrm{Ex}}caligraphic_H start_POSTSUPERSCRIPT roman_Ex end_POSTSUPERSCRIPT.c.The previousk𝑘kitalic_kcycles in lifelong learning and be folded to separate the future from the past. | 2502.00197v1 | 4 |
|
Figure 1:Overview of the synthetic data creation process forDermaSynth. State of the art large language model (Gemini 2.0) were used to generate synthetic and clinically relevant image-text pairs[10]. | 2502.00196v1 | 1 |
|
Figure 1:Illustration ofRAP-GNN. The pretrained embeddingfϕpresubscriptsuperscript𝑓preitalic-ϕf^{\text{pre}}_{\phi}italic_f start_POSTSUPERSCRIPT pre end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_ϕ end_POSTSUBSCRIPTis frozen, while the classifiercθsubscript𝑐𝜃c_{\theta}italic_c start_POSTSUBSCRIPT italic_θ end_POSTSUBSCRIPTis optimized. All GNN weights,𝐰(l),l=1,…,Lformulae-sequencesuperscript𝐰𝑙𝑙1…𝐿\mathbf{w}^{(l)},\ l=1,\ldots,Lbold_w start_POSTSUPERSCRIPT ( italic_l ) end_POSTSUPERSCRIPT , italic_l = 1 , … , italic_Lare diagonal matrices, randomly sampled on-the-fly in each forward pass. | 2502.00190v1 | 1 |
|
Figure 2:Training time (a), inference time (b), and accuracy (c) for node classification on PubMed with GCN backbone. The runtimes gap between ourRAP-GNNand End-to-End widens as the number of layers increases, while obtaining superior accuracy than End-to-End. Accuracy results in (c) validateObservation2, showingRAP-GNNmaintains stability in deep architectures, unlike Random Weights. | 2502.00190v1 | 2 |
|
Figure 3:GPU memory usage comparison on on the PubMed dataset:RAP-GNNrequiring only a third of the memory compared to the End-to-End (GCN backbone). | 2502.00190v1 | 3 |
|
Figure 4:Comparison of the mean ofVar(𝕙(l))Varsuperscript𝕙𝑙\text{Var}\big{(}\mathbb{h}^{(l)}\big{)}Var ( blackboard_h start_POSTSUPERSCRIPT ( italic_l ) end_POSTSUPERSCRIPT )(in log scale) across layers for End-to-End andRAP-GNNon the PubMed dataset using the same GCN architecture with residual connections withL=64𝐿64L=64italic_L = 64andd=256𝑑256d=256italic_d = 256.RAP-GNNincreases embedding variance with higher rate compared to End-to-End as the number of layers grows, consistent with the theoretical result inObservation2. | 2502.00190v1 | 4 |
|
Figure 6:Comparison of GPU memory consumption on the ogbn-proteins dataset. OurRAP-GNNsignificantly reduces GPU memory consumption. | 2502.00190v1 | 8 |
Subsets and Splits