L-WISE: Boosting Human Visual Category Learning Through Model-Based Image Selection and Enhancement

Gabriel Kreiman; Guy Gaziv; James J. DiCarlo; Morgan B. Talbot

REVIEW 5 major objections 7 minor 1 cited by

L-WISE: Boosting Human Visual Category Learning Through Model-Based Image Selection and Enhancement

T0 review · 5 major / 7 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Model-chosen, model-enhanced training images lift human visual category learning by 33-72%.

desk verdict L-WISE is a genuine new result—robustified ANN logits both predict human difficulty and enhance human recognition—but the trial-level chi-square tests make the headline gains less certain than reported. read the letter →

arxiv 2412.09765 v4 pith:4ZFJRK6Y submitted 2024-12-12 cs.CV cs.HC

Morgan B. Talbot , Gabriel Kreiman , James J. DiCarlo , Guy Gaziv This is my paper · ORCID

classification cs.CVcs.HC

keywords visualcategorylearningmachineteachingimageenhancementhuman-AIalignmentperceptualcurriculumdermoscopyhistology

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to show that a computer vision model hardened against adversarial attacks can be used as a teaching aid for human visual learning. The model's confidence score for the correct category serves two roles: it predicts how hard each image is for a human, and it provides a gradient direction for altering the image so the correct category is easier to see. Combining those roles into a training curriculum — show easier, model-enhanced images first and then fade both aids — improves novice humans' accuracy on unmodified test images across three tasks: moths, skin lesions, and colon histology. The reported gains are 33-72% on the accuracy margin above chance, plus a 20-23% shortening of training time with the same number of trials. If correct, this is a route to faster training in visually demanding specialties such as pathology and dermatology.

What carries the argument

L-WISE's engine is the ground truth logit $L_{\mathrm{gt}}$: the pre-softmax activation a robustified ResNet-50 model assigns to an image's correct category. That scalar is used as a difficulty percentile for image selection, with the allowed difficulty ramping up stepwise during training, and as the objective of projected gradient ascent in pixel space, which perturbs the image within an $\ell_2$ budget $\epsilon$ that tapers exponentially from 8 to 0. A variant of the optimization also subtracts competing-class logits to sharpen the category signal. The machinery works because the model is adversarially trained, which aligns its gradients with human perception; vanilla models do not produce the effect.

What would settle it

A decisive check is a reanalysis that respects grouping in the data: fit mixed-effects logistic regressions on trial-level responses with random intercepts for participant and image, or average accuracy within participant and test the group difference with a participant-level test. If the histology effect (reported p = 0.023) and the moth effect (p < 0.001) do not survive participant-level clustering, or if an intent-to-treat analysis that includes the higher control-group dropout rate removes the gain, the learning-boost claim is an artifact of pseudoreplication.

Watch

Extended reading notes

Core claim

Central claim: robustified ANNs, used out of the box, provide a single number — the logit of the ground truth class — that is both the best available predictor of image difficulty for humans and a valid objective for making images easier for humans to classify. The paper then claims that scheduling these two operations together (L-WISE) yields a learning boost: on a 4-way moth classification task the average test-time accuracy margin above chance rose 57.6% (p < 0.001); on a 4-way dermoscopy task it rose 72.3% (p < 0.001); on a binary colon histology task it rose 33.1% (p = 0.023). In the same experiments, training time fell by 20-23% at a fixed trial count, and both mechanisms contributed: enhancement tapering alone and difficulty selection alone each raised accuracy, with full L-WISE numerically but not always significantly better.

Load-bearing premise

The load-bearing premise is that the trial-level statistics are valid: the gains are computed with chi-square tests on pooled trials treated as independent, the difficulty-prediction AUC is estimated with cross-validation where the same images recur across folds, and control participants withdrew at higher rates, so if any of these inflate the comparisons, the learning boost shrinks or disappears.

Editorial extensions

If this is right

Human novices trained with L-WISE reach higher accuracy on unmodified, randomly selected test images than controls trained on the same trial count without model assistance.
The same number of training trials takes 20-23% less time when L-WISE is active, meaning learning speed improves without extra exposure.
The ground truth logit of a robustified network is a better human-difficulty estimator than c-score, prediction depth, or adversarial robustness, so it can be used to rank images for any future curriculum.
Standard image-enhancement tools (CLAHE, multi-scale Retinex, Lightroom auto) do not raise human accuracy; the effect is specific to category-directed, model-guided enhancement.
Both enhancement tapering and difficulty selection contribute, but the benefit of enhancement depends on easy-to-hard ordering while difficulty selection's benefit mostly comes from showing easier images on average.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper, the difficulty signal could drive fully adaptive curricula that adjust selection and enhancement online per learner; the paper only tested fixed schedules.
If enhancement works by amplifying category-defining features, such as moth wing spots, L-WISE-trained learners may be vulnerable on test images lacking those features; a targeted test would withhold the highlighted feature and compare generalization.
The method's clinical promise is tied to dataset representativeness: enhancement and selection drawing from a light-skinned dermoscopy set could teach learners a biased 'average' melanoma, as the paper acknowledges.
Because logit-maximization sometimes hallucinates features, such as an extra antelope, a follow-up could separate learners who saw hallucinated features from those who did not and measure whether hallucinations help or mislead learning.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: this is a real new result, not just an increment. The idea of inverting the Gaziv et al. (2023) perturbation direction—from disrupting percepts to enhancing them—and applying it to human curriculum learning is genuinely fresh. They show robustified-ANN ground-truth logits predict human error rates better than prior c-score/prediction-depth methods (AUC ~0.72 in their comparison), and logit-maximizing perturbations make images easier for humans to recognize. The three-task replication (moths, dermoscopy, histology) with both image selection and enhancement tapering is a solid proof-of-concept. The ablations (ET, DS, shuffled) are thoughtful and give a real decomposition of the effects. They also ship code and state a reproducibility protocol, which is more than many papers in this area do.

The soft spots: the headline statistics rest on chi-square tests that pool ~30 participants × 40 trials per group and ignore participant clustering. The histology effect (p=0.023) is small and could easily lose significance in a mixed-effects or participant-level analysis. The other two effects (p<0.001) look robust in magnitude, so this is a breadth-of-claim problem more than a core-effect problem. The AUC for difficulty prediction uses trial-level cross-validation where the same image likely appears in both train and test folds, so the 0.72 is probably slightly optimistic. The differential dropout in Appendix S9 (controls more likely to withdraw) is real and threatens random assignment—though to their credit the authors report it plainly and it does not by itself explain the accuracy gains.

These don't sink the paper, but they do mean the effect sizes are upper-bound estimates. A revision with a clustered analysis would settle it. The limitations section is unusually honest about homogenizing enhancements and dataset bias, especially the HAM10000 skin-tone concerns.

Who should read it: anyone working on model-human alignment, perceptual learning, or machine teaching. It is a useful demonstration that robustified models are behaviorally aligned enough to guide human learning. Would I send it to reviewers? Yes—it deserves serious refereeing, and the weaknesses are addressable with a bit more statistical work.

Referee Report

5 major / 7 minor

Summary. The paper proposes L-WISE (Logit-Weighted Image Selection and Enhancement), a method that uses adversarially robustified ANNs to (1) predict the human-perceived difficulty of an image via the model's ground-truth logit activation and (2) generate image perturbations that maximize that logit, thereby making the ground-truth category more perceptible. These two abilities are combined into a training-phase curriculum for human learners: images are selected from a growing allowed-difficulty percentile and enhanced with a shrinking pixel budget over successive blocks. The authors report that across three image domains (moth photographs, dermoscopy images, and histology images), L-WISE increases test-time accuracy margin above chance by 57.6%, 72.3%, and 33.1%, respectively, and reduces training duration by 20–23%, while keeping the number of training trials constant. They also claim that the ground-truth logit from a robustified ResNet-50 is a state-of-the-art predictor of human trial-level accuracy (AUC=0.72), and they include ablation experiments separating the contributions of enhancement tapering and difficulty selection.

Significance. If the central claims hold, this is a genuinely novel and potentially impactful application of perceptually aligned ANN gradients to human learning, with clear connections to machine teaching, perceptual learning, and clinical training. The paper's strengths include the use of three independent participant groups across three diverse image domains, the inclusion of two ablation conditions with shuffled controls, the explicit attempt to compare against prior difficulty predictors (c-score, prediction depth, adversarial robustness), and a reproducibility statement promising released code and a web-based experimental framework. The scientific claim that a fixed, pre-trained robustified network can both rank images by human difficulty and generate perturbations that improve human recognition is well supported by the pattern of results if the statistical tests are valid.

major comments (5)

[§4.2, Fig. 4A] The three headline chi-square tests (p<0.001, p<0.001, p=0.023) pool all test-phase trials across participants (approximately 30 participants per group x 40 trials each) and treat each trial as independent. Responses from the same participant are correlated, and the same test images may also recur across participants, so the effective sample size is far smaller than the ~1200 trials per group. The histology result (p=0.023) is the most fragile: a small absolute accuracy gain on a binary task could easily lose significance in a participant-level reanalysis. Please provide a clustered analysis, e.g., a mixed-effects logistic regression with random intercepts for participant (and, if images are reused, for image), or a non-parametric comparison of per-participant mean test accuracies. This is load-bearing because the abstract's '33-72%' gains depend on all three tasks being significant.
[§4.1, Fig. 1A] The AUC of 0.72 is reported 'under 10-fold cross-validation' on trial-level responses, but the text does not state how the folds were split. If folds were split by trial rather than by image or by participant, the same image (and the same participant) can appear in both training and test folds, inflating the AUC via image-specific memorization. Please clarify the cross-validation scheme and, if the split is currently trial-level, re-run the analysis with image-level and participant-level splits. The same issue applies to Appendix Fig. S6, where AUCs are compared across predictors under 500-fold cross-validation.
[Appendix S9] The differential dropout analysis shows that control participants withdrew at significantly higher rates in the moth task (p=0.02) and dermoscopy task (p=0.041). The authors report this as a property of the intervention, but it also threatens the internal validity of the group comparisons: if withdrawal is related to task performance, the completers in the control group may be a non-random subset, biasing the estimated control mean in an unknown direction. Please add a sensitivity analysis (e.g., inverse-probability weighting, worst-case replacement, or at least a comparison of early-task accuracy between eventual completers and dropouts) and discuss how the primary conclusions might change.
[Figure 4A and Table 1] Training duration reductions of 20% (moths), 23% (dermoscopy), and 22% (histology) are reported without any accompanying statistical test. The figure caption says 'All differences ... are statistically significant (χ2(1) test, p<0.05)', but a chi-square test is not appropriate for duration, which is a continuous variable, and no duration p-values or test statistics are given anywhere in the text. Please provide appropriate tests for the duration comparisons (e.g., Welch's t-test or Mann-Whitney U on per-participant durations), and state whether the duration advantage is significant in each task.
[Appendix S6, Fig. S6] The claim that the ground-truth logit is 'a more accurate predictor of image difficulty for humans than prior methods' should be supported by a comparison that is strictly controlled. Fig. S6 reports AUCs for several single features and combinations, but the number of folds (500) and the exact feature definitions (e.g., how c-score and prediction depth are computed on the restricted 16-class set) are not fully specified, and it is unclear whether the cross-validation folds are trial-level or image-level. If the comparison uses different cross-validation procedures, the AUC difference may be an artifact. Please specify the fold construction and report confidence intervals for the differences.

minor comments (7)

[§4.1] The text says the logistic regression 'pooled responses to original images with those to modified control-group images' but does not define what 'modified control-group images' are until later; specify that these are the unmodified images in the control group of the learning experiments, or clarify in the earlier sentence.
[Equation (1) and Appendix S1] The constraint in Eq. (1) is written as ||δ|| < ε, but the projection in Eq. (2) is onto a hypersphere of radius ε. Clarify whether the constraint is the non-strict inequality ||δ|| ≤ ε, and state the step size and number of iterations for the reported experiments.
[§3.4 and Fig. 2F] The text refers to an 'exponential tapering' of ε, but the Methods describe halving ε per block. Halving is geometric (exponential) in block index, but the phrase should be defined explicitly; also state the starting ε and number of blocks for the histology task.
[Abstract and §4.2] The abstract reports 'accuracy gains of 33-72%' while the text reports gains in 'accuracy margin above chance levels'. Please make the basis of the percentage consistently explicit (margin above chance vs. absolute accuracy) to avoid overstatement.
[Table 1] The table lists 'ET (shuffled)' and 'DS (shuffled)' without a footnote explaining the shuffling procedure; add a note that shuffling flattens the easy-to-hard trend, as defined in the main text.
[Appendix S9] The binomial test equation uses p for the probability of assignment to control, but the text does not report the actual p used; state the assignment probabilities across the experimental conditions.
[Reproducibility Statement] The paper refers to a 'Code Webpage' but no URL is given; include a persistent link or repository identifier.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: L-WISE's difficulty score and enhancement are model outputs validated on external human data; the learning gains are measured on unmodified held-out images.

full rationale

The derivation chain is empirical and self-contained. The difficulty score is the raw ground-truth logit Lgt of a fixed robustified ANN (Section 3.2); its link to human accuracy is established by fitting a logistic regression to new human trial responses (Section 4.1, Fig. 1A, AUC=0.72) and per-domain control data (Fig. S14). The enhancement is a projected gradient ascent that maximizes the same Lgt (Eq. 1); its effect on human accuracy is measured directly on human participants (Fig. 1B, Fig. S9), not inferred from the logit increase. L-WISE combines these two model-based operations into a curriculum, and the outcome is test accuracy on unmodified, randomly selected held-out images from participants randomized to L-WISE or control (Section 4.2). No step defines the target quantity (human learning gain) in terms of the model quantity used to generate the intervention: the model is never fit to human learning outcomes. Self-citations to Gaziv et al. (2023) supply the adversarial training recipe and prior evidence of perceptually aligned gradients, but the present paper independently re-tests the enhancement and difficulty predictions with new human data, so those citations are not load-bearing. The statistical concerns raised by pooling trials across participants and by potential image reuse in the AUC cross-validation are validity limitations, not circularity.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a handful of hand-selected schedules and robustness levels (listed as free parameters) and on several domain assumptions about human behavior and dataset labels. No new physical or formal entities are introduced; the method combines existing models and perturbation techniques.

free parameters (5)

Adversarial training epsilon for guide models = 3.0 (ImageNet), 1.0 (iNaturalist and all fine-tuning)
Degree of robustification determines perceptually aligned gradients. Inherited from Gaziv et al. (2023) for ImageNet; lowered to 1.0 because the models could not learn the fine-grained tasks at 3. Not fitted to human outcomes.
L-WISE enhancement schedule = epsilon = 8 in block 1, halved per block to 0 after block 6
Hand-selected after a pilot with epsilon = 20 produced poor generalization; fixed for all three learning tasks.
L-WISE difficulty schedule = d1 = 0.1, +0.15 per block to 1.0 at blocks 7-8
A priori choice of easy-to-hard ramping; not optimized in this proof-of-concept.
Competing-class suppression weight alpha = 0 for basic animals, 1 for fine-grained tasks
For fine-grained tasks, minimizing competing-class logits improved subjective enhancement quality; chosen by the authors.
Projected gradient ascent step size and step count = eta = 0.5, ceil(2*epsilon) steps (ceil(4*epsilon) for XCiT)
Implementation detail; adjusted for the guide model type.

assumptions (5)

domain assumption The ground-truth logit of a robustified ANN is a valid monotonic proxy for human recognition difficulty in the tested domains.
The foundation of both L-WISE mechanisms; validated retrospectively on human data (Fig. 1A, Fig. S14), but assumed when designing the curricula.
domain assumption Trial-level responses can be treated as independent for chi-square significance tests.
Invoked in Section 4.2 and Fig. 4A; doubtful because each participant contributes 40 test trials and repeated images are shown to many participants.
domain assumption The 17ms screening and warm-up procedure ensures participants perform the near-threshold animal task as a perceptual categorization task.
Participants had to pass a 200ms screening with >=24/32 correct; below-threshold guesses could otherwise dominate.
domain assumption Dataset ground-truth labels are trustworthy; MHIST uses majority vote of 7 pathologists.
Relied on when computing Lgt and when scoring human accuracy; the paper acknowledges annotator disagreement in Fig. S13C.
domain assumption Attention-check accuracy >= 90% filters out inattentive participants.
Standard behavioral filtering; 10-12.5% of trials were circle/triangle checks.

how reviews work

0 comments

Cite this review

Pith. "Pith review of L-WISE: Boosting Human Visual Category Learning Through Model-Based Image Selection and Enhancement." pith.science (2026). https://pith.science/paper/4ZFJRK6Y

@misc{pith2026241209765,
  author       = {Pith},
  title        = {Pith review of: L-WISE: Boosting Human Visual Category Learning Through Model-Based Image Selection and Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4ZFJRK6Y}},
  note         = {Machine review of arXiv:2412.09765}
}

read the original abstract

The currently leading artificial neural network models of the visual ventral stream - which are derived from a combination of performance optimization and robustification methods - have demonstrated a remarkable degree of behavioral alignment with humans on visual categorization tasks. We show that image perturbations generated by these models can enhance the ability of humans to accurately report the ground truth class. Furthermore, we find that the same models can also be used out-of-the-box to predict the proportion of correct human responses to individual images, providing a simple, human-aligned estimator of the relative difficulty of each image. Motivated by these observations, we propose to augment visual learning in humans in a way that improves human categorization accuracy at test time. Our learning augmentation approach consists of (i) selecting images based on their model-estimated recognition difficulty, and (ii) applying image perturbations that aid recognition for novice learners. We find that combining these model-based strategies leads to categorization accuracy gains of 33-72% relative to control subjects without these interventions, on unmodified, randomly selected held-out test images. Beyond the accuracy gain, the training time for the augmented learning group was also shortened by 20-23%, despite both groups completing the same number of training trials. We demonstrate the efficacy of our approach in a fine-grained categorization task with natural images, as well as two tasks in clinically relevant image domains - histology and dermoscopy - where visual learning is notoriously challenging. To the best of our knowledge, our work is the first application of artificial neural networks to increase visual learning performance in humans by enhancing category-specific image features.

Figures

Figures reproduced from arXiv: 2412.09765 by the authors.

**Figure 1.** Robustified ANNs can be used out-of-the-box as image recognition difficulty estimators and ground truth percept enhancers. We consider a 16-way basic animal classification task. Panel A1 shows the correspondence between human categorization accuracy and model-computed ground truth logit activation values. The curve denotes a logistic regression model predicting the probability of a correct response using only the l… view at source ↗

**Figure 2.** Robustified ANNs can be used to boost image category learning in humans. A novice human learner undertakes a challenging image categorization task, which consists of a training phase (B) and a test phase (C). Images for both phases are randomly drawn from a labeled image dataset of unfamiliar finegrained categories (A). Feedback (correct/incorrect, with indication of the correct category) is delivered after each tr… view at source ↗

**Figure 3.** Novice learners who had their curriculum augmented by our method showed improved test-time categorization accuracy for previously unfamiliar categories. This figure shows empirical results from a 4-way fine-grained moth species classification task. Panel A shows examples of the 4 moth classes, side-by-side with their model-enhanced versions at the highest pixel budget used in our experiments (ϵ = 8). While subtle,… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Our approach can boost time efficiency and final accuracy of image category learning for humans across varied image domains, including in clinically relevant tasks. Panel A compares the mean test-phase accuracy and training-phase duration of human participants who wer…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Noninvasive precision modulation of high-level neural population activity via natural vision perturbations
q-bio.NC 2025-06 conditional novelty 7.0 of 10

Model-designed image perturbations applied to natural images produce targeted, measurable shifts in macaque inferior temporal cortex population activity, with intended modulation several times larger than orthogonal s...

Reference graph

Works this paper leans on

20 extracted references · 19 canonical work pages · cited by 1 Pith paper

[1]

26 Published as a conference paper at ICLR 2025 0 4 8Enhance ϵ 0 40 80Diff

were less likely to withdraw. 26 Published as a conference paper at ICLR 2025 0 4 8Enhance ϵ 0 40 80Diff. %ile 0 20 40 60 80 100 120 140 160 0.2 0.3 0.4 0.5 0.6 0.7 0.8Accuracy (original groundtruth) Control ϵ (pixel budget) L-WISE ϵ (pixel budget) Control DifficultyL-WISE Difficulty Control accuracy L-WISE accuracy (Ours) Train (feedback) Test (no feedba...

work page 2025
[3]

ImageNet: A large-scale hier- archical image database

Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hier- archical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pp. 248–255. IEEE,

2009
[5]

Amicable aid: Perturbing images to improve classification performance

Juyeop Kim, Jun-Ho Choi, Soobeom Jang, and Jong-Seok Lee. Amicable aid: Perturbing images to improve classification performance. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1–5. IEEE,

work page 2023
[7]

iNat Challenge 2021 - FGVC8,

14 Published as a conference paper at ICLR 2025 Grant Van Horn and Oisin Mac Aodha. iNat Challenge 2021 - FGVC8,

work page 2025
[9]

Gradient-based algorithms for machine teach- ing

Pei Wang, Kabir Nagrecha, and Nuno Vasconcelos. Gradient-based algorithms for machine teach- ing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 1387–1396, 2021a. Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4555–4576,...

work page 2021
[10]

The results of this experiment are summarized in Fig. S11A. We hypothesized that logit-based enhancement would provide superior results, particularly for images that started off with low cross-entropy loss. We further hypothe- sized that enhancements would be less effective for training images due to overfitting of the guide model on them. The results sho...

work page 2015
[13]

rodent” class (as many people may not recognize these as rodents), and we did not include eels in the “fish

Regions that were changed more on average are more red, and regions that were changed less on average are more blue. The heat map was generated by averaging the normalized absolute pixel value changes across all 2400 ImageNet validation set images that we used for our 16-way animal classification experiments. S3 D ETAILS ON 16- WAY IMAGE NET ANIMAL CATEGO...

work page 2008
[14]

RN50 = ResNet-50, and ϵ values in the labels for each bar show the magnitude of the adversarial perturbations during adversarial training

using gradients from the corresponding guide model. RN50 = ResNet-50, and ϵ values in the labels for each bar show the magnitude of the adversarial perturbations during adversarial training. lowed by cropping/padding), color-jittered, and randomly cut-out “views” of each image, a strategy inspired by Ganz & Elad (2024) that extends DiffAugment (Zhao et al...

work page 2024

Show all 20 references

[16]

confident

Panel C of Fig. S13 shows the rela- tionship between the ground truth logit from robust ResNet-50 model and how many of the 7 expert annotators of the MHIST histology dataset Wei et al. (2021) agreed on the same category label. On average, the model is more “confident” in its ...

2021
[18]

enhancement taper

There is a statistically significant difference between the test-phase performance of the L-WISE participants and that of the control participants ( χ2(1) test, p <0.001) in panel G but not for the pilot experiment in panel M. Notably, the last portion of the training phase do...

2025
[19]

Enhancement Taper (shuffled)

= 1− d−1X k=0 n k pk (1 − p)n−k (3) Similarly, in the dermoscopy category learning experiment, 13 participants withdrew, of whom 6 were from the control group. In this case, Equation 3 evaluates to a probability of p = 0.041, again indicating that participants in the control g...

2023
[20]

grasshopper

The third row shows a 5x magnified version of the difference between the enhanced image and the original, and the bot- tom row shows a heat map where red regions correspond to larger changes and blue regions correspond to smaller changes. S1 D ETAILS ON TRAINING AND USING ROBU...

2023
[21]

Leda,” “Ajax,

participants to be roughly 10%, with a standard deviation in accuracy of roughly 10% in each group. H0 : µ1 − µ2 = 0 H1 : µ1 − µ2 ̸= 0 Given: δ = 0.1 (estimated mean difference) σ = 0.1 (estimated standard deviation) α = 0.05 (significance level) 1 − β = 0.8 (power) Estimated ...

2025
[224]

Any single-channel grayscale images were converted to RGB before further processing. The baseline enhancement algorithms Contrast-Limited Adaptive Histogram Equalization (CLAHE (Zuiderveld, 1994)), Multi-Scale Retinex with Color Restoration (MSRCR (Jobson et al., 1997; Petro e...

1994
[1994]

The top row shows original ImageNet images, and the second row shows the same images after enhancement by robustified ResNet-50 (training ϵ =

15 Published as a conference paper at ICLR 2025 Appendix CrabFish BirdLizard DogInsectMonkey Original images Enhanced with pixel budget ϵ = 20 Difference heat map Difference (magnified 5x) Figure S1: Ground truth logit enhancement with robustified ANNs leads to semantically me...

2025
[2015]

A light recipe to train robust vision transformers

Edoardo Debenedetti, Vikash Sehwag, and Prateek Mittal. A light recipe to train robust vision transformers. In 2023 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML), pp. 225–253. IEEE,

2023
[2019]

12 Published as a conference paper at ICLR 2025 Thomas Fel, Ivan Felipe, Drew Linsley, and Thomas Serre

URL https://github.com/MadryLab/robustness. 12 Published as a conference paper at ICLR 2025 Thomas Fel, Ivan Felipe, Drew Linsley, and Thomas Serre. Harmonizing the object recognition strategies of deep neural networks with humans. In Advances in Neural Information Processing ...

2025
[2021]

Grant Van Horn, Elijah Cole, Sara Beery, Kimberly Wilber, Serge Belongie, and Oisin Mac Aodha

URL https: //kaggle.com/competitions/inaturalist-2021. Grant Van Horn, Elijah Cole, Sara Beery, Kimberly Wilber, Serge Belongie, and Oisin Mac Aodha. Benchmarking representation learning for natural world image collections. In Proceedings of the IEEE/CVF Conference on Computer...

2021
[2022]

Teaching categories to human learners with visual explanations

13 Published as a conference paper at ICLR 2025 Oisin Mac Aodha, Shihan Su, Yuxin Chen, Pietro Perona, and Yisong Yue. Teaching categories to human learners with visual explanations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3820–3828,

2025
[2023]

Cormier, Yan Xing, Meichun Ding, Jeffrey E

Janice N. Cormier, Yan Xing, Meichun Ding, Jeffrey E. Lee, Paul F. Mansfield, Jeffrey E. Gershen- wald, Merrick I. Ross, and Xianglin L. Du. Ethnic differences among patients with cutaneous melanoma. Archives of Internal Medicine, 166(17):1907–1914,

1907

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

26 Published as a conference paper at ICLR 2025 0 4 8Enhance ϵ 0 40 80Diff

were less likely to withdraw. 26 Published as a conference paper at ICLR 2025 0 4 8Enhance ϵ 0 40 80Diff. %ile 0 20 40 60 80 100 120 140 160 0.2 0.3 0.4 0.5 0.6 0.7 0.8Accuracy (original groundtruth) Control ϵ (pixel budget) L-WISE ϵ (pixel budget) Control DifficultyL-WISE Difficulty Control accuracy L-WISE accuracy (Ours) Train (feedback) Test (no feedba...

work page 2025

[2] [3]

ImageNet: A large-scale hier- archical image database

Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hier- archical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pp. 248–255. IEEE,

2009

[3] [5]

Amicable aid: Perturbing images to improve classification performance

Juyeop Kim, Jun-Ho Choi, Soobeom Jang, and Jong-Seok Lee. Amicable aid: Perturbing images to improve classification performance. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1–5. IEEE,

work page 2023

[4] [7]

iNat Challenge 2021 - FGVC8,

14 Published as a conference paper at ICLR 2025 Grant Van Horn and Oisin Mac Aodha. iNat Challenge 2021 - FGVC8,

work page 2025

[5] [9]

Gradient-based algorithms for machine teach- ing

Pei Wang, Kabir Nagrecha, and Nuno Vasconcelos. Gradient-based algorithms for machine teach- ing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 1387–1396, 2021a. Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4555–4576,...

work page 2021

[6] [10]

The results of this experiment are summarized in Fig. S11A. We hypothesized that logit-based enhancement would provide superior results, particularly for images that started off with low cross-entropy loss. We further hypothe- sized that enhancements would be less effective for training images due to overfitting of the guide model on them. The results sho...

work page 2015

[7] [13]

rodent” class (as many people may not recognize these as rodents), and we did not include eels in the “fish

Regions that were changed more on average are more red, and regions that were changed less on average are more blue. The heat map was generated by averaging the normalized absolute pixel value changes across all 2400 ImageNet validation set images that we used for our 16-way animal classification experiments. S3 D ETAILS ON 16- WAY IMAGE NET ANIMAL CATEGO...

work page 2008

[8] [14]

RN50 = ResNet-50, and ϵ values in the labels for each bar show the magnitude of the adversarial perturbations during adversarial training

using gradients from the corresponding guide model. RN50 = ResNet-50, and ϵ values in the labels for each bar show the magnitude of the adversarial perturbations during adversarial training. lowed by cropping/padding), color-jittered, and randomly cut-out “views” of each image, a strategy inspired by Ganz & Elad (2024) that extends DiffAugment (Zhao et al...

work page 2024