Pith. sign in

REVIEW 4 major objections 6 minor 68 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

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

T0 review · grok-4.5

Autistic–neurotypical emotion-judgment differences are sparse at the image level, and behavior-aligned neural nets can both find and transform the faces that reveal them.

2026-07-10 05:48 UTC pith:TLDVMFRE

load-bearing objection Solid methods paper: sparsity and CLIP-guided selection are real; synthesis is weaker and the Methods/Results loss descriptions do not match. the 4 major comments →

arxiv 2607.08533 v1 pith:TLDVMFRE submitted 2026-07-09 cs.AI cs.LG

AI-guided stimuli discovery and generation to optimize facial emotion perception studies in autism

Kushin Mukherjee , Na Yeon Kim , Maren Wehrheim , Ralph Adolphs , Kohitij Kar This is my paper
classification cs.AI cs.LG
keywords autismfacial emotion perceptionstimulus selectionartificial neural networksgenerative adversarial networksbehavioral phenotypingimage-level diagnosticity
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Facial emotion tasks often produce weak or inconsistent autistic–neurotypical group differences because those differences are not spread evenly across faces: a few diagnostic expressions carry most of the separation, and averaging over the rest dilutes the signal. This paper trains separate artificial neural network readouts on autistic and neurotypical image-level judgments, then uses the predicted gap between those readouts to rank new faces before anyone is tested. In an independent cohort, faces chosen this way produced larger group separation than matched random faces, with the best model clearly beating random sampling. The same models were then coupled to a generative face model to edit already-diagnostic images toward predicted agreement; under phenotype-matched validation, the edited faces reduced the measured gap relative to their originals. The broader claim is that behavioral phenotyping can stop treating stimulus sets as fixed background and instead use image-computable models to discover and perturb the conditions under which perception diverges or converges.

Core claim

Autistic–neurotypical differences in facial emotion judgments are concentrated in a sparse subset of high-leverage images rather than expressed uniformly. Population-specific ANN models that predict those image-level judgments can prospectively select novel faces that produce larger group separation than random sampling in a new cohort, and the same models can guide generative transformations of diagnostic faces that reduce separation when validation participants are matched to the targeted response phenotype.

What carries the argument

Population-specific ridge-regression readouts on fixed ANN visual embeddings, whose predicted autistic–neurotypical gap both ranks candidate faces for selection and supplies the loss for closed-loop GAN latent-code optimization that synthesizes gap-reduced expressions.

Load-bearing premise

The synthesis result assumes that matching new participants by correlation to the original group response templates is a fair test of gap reduction rather than mainly recovering the subspace the optimizer was trained to fix.

What would settle it

In a new independent cohort, without phenotype matching, CLIP-ranked faces fail to beat identity- and intensity-matched random faces on |ASD–NT| happy-response difference, or the synthesized faces fail to reduce that gap relative to their diagnostic bases under the same leave-one-image-out protocol.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper argues that autistic–neurotypical differences in facial emotion judgments are sparse at the image level rather than uniform across stimuli, and that population-specific ANN readouts on shared visual features can both (i) prospectively select novel faces that enlarge group separation and (ii) guide GANmut transformations that reduce separation. Reanalysis of Wang & Adolphs (2017) shows diagnostic images form a high-leverage tail. Models trained on that dataset are applied to MSFDE; in an independent lab cohort (12 ASD, 13 NT), ANN-selected sets vary by architecture, with CLIP clearly outperforming matched random sets (|ASD–NT| 0.149 vs 0.091, empirical p=.006), and selection success tracking NT behavioral alignment (ρ=.82). Closed-loop synthesis then optimizes GANmut latent codes to reduce predicted group divergence; under leave-one-image-out correlation-based phenotype matching in a Prolific cohort, mean gap falls from 0.138 to 0.076 (t(14)=2.38, one-tailed p=.016). AU-region summaries describe structured but multi-region facial changes.

Significance. If the dual claim holds, the work supplies a concrete template for moving autism behavioral assays from stimulus-averaged summaries to image-computable, prospectively optimized stimulus design—an important methodological shift for heterogeneous neurodevelopmental phenotypes. Strengths include prospective testing of selection on a new stimulus set and independent lab cohort, explicit architecture comparison rather than a single black-box model, and a closed-loop synthesis intervention that goes beyond post-hoc explanation. The sparsity result (Figs. 1B, 3E) is a useful reframing of why facial-emotion findings have been inconsistent. The synthesis half, if cleaned up, would be a stronger validity test than selection alone. Code and de-identified data are promised, which supports reproducibility.

major comments (4)
  1. Results (Fig. 3B) report that the average selected-versus-random uplift across seven ANNs is not significant (mean 0.010 ± 0.012 SEM; paired t(6)=0.84; one-tailed p=.217); only CLIP clearly beats random (empirical p=.006). The abstract and strongest framing present “model-selected images produced larger behavioral differences than matched random images” as a general result. That claim should be restated as architecture-dependent, with CLIP (and NT-alignment) as the operative finding, or the multi-model average should be demoted from a positive result.
  2. Methods vs Results disagree on the synthesis objective. Results and Fig. 4A describe minimizing the predicted autistic–neurotypical difference on the candidate image (Δ(θ,ρ) between population-specific “happy” scores). Methods define the loss as (neurotypical score of the original image − autistic score of the synthesized image)². These are not the same quantity. If Methods is correct, the optimizer does not implement the closed-loop intervention claimed in Results, so the behavioral gap reduction is not a clean test of that intervention. This inconsistency must be resolved with the actual loss used, and any mismatch between claimed and implemented objective should be stated.
  3. Closed-loop synthesis validation (Results §Closed-loop synthesis; Fig. 4C; Methods) uses leave-one-image-out correlation matching of Prolific participants to the original Wang & Adolphs group response templates, then measures gap reduction on held-out pairs. Discussion correctly scopes this as testing attenuation within the targeted image-level phenotype, but the abstract and primary results presentation still report phenotype-matched reduction as the synthesis result without that qualifier. The claim should be limited to the matched subspace, and unmatched or randomly sampled cohort analyses (even if null or weaker) should be reported so readers can judge generalizability.
  4. Lab selection cohort is small (n=12 ASD, 13 NT; Table S1) and the online synthesis cohort relies on self-reported diagnosis plus SRS/AQ (Fig. S3). Trait-binned analyses (Fig. S4) help, but the synthesis effect size and selection uplift need either larger independent replication or explicit power/reliability bounds before the framework is positioned as ready for assay construction. At minimum, report image-level reliability and subject-level stability of |ASD–NT| for selected vs random sets.
minor comments (6)
  1. Figure 2C hypotheses (H0/H1/H2) are useful but the main text should state which architectures support H2 before the multi-model average is discussed.
  2. Clarify ridge regularization, cross-validation folds, and whether decoder hyperparameters were chosen on Wang & Adolphs only or tuned with any MSFDE information.
  3. Fig. 3C uses ΔP(happy)=Control−ASD while elsewhere Δ is autistic−NT; keep signed conventions consistent.
  4. AU analysis (Fig. 5) is appropriately descriptive; state explicitly that AUCanvas masks come from neutral references and that pixel-thresholding for overlays is not used in the quantitative vectors.
  5. Abstract says “independent cohort” for selection and “phenotype-matched validation” for synthesis; keep that distinction in the Results lead sentences as well.
  6. Minor typos: “mechani stic”, “netw ork”, spacing artifacts in the abstract/intro PDF text.

Circularity Check

1 steps flagged

Prospective selection is independent; synthesis success is measured only inside leave-one-out phenotype-matched subspace the optimizer targeted, not forced by construction.

specific steps
  1. fitted input called prediction [Results §Closed-loop synthesis; Fig. 4C; Methods (phenotype-matched validation)]
    "For each held-out image pair, neurotypical participants were selected according to the correlation between their responses to the remaining base images and the original neurotypical behavioral template. Autistic participants were selected in the same way, using the original autistic behavioral template. ... Under this correlation-based phenotype-matched validation, synthesized images reliably reduced autistic–neurotypical behavioral divergence (Figure 4C)."

    Participant inclusion is conditioned on matching the same group-specific image-level response templates that defined the synthesis objective. Gap reduction is then reported only inside that reselected subspace. This is not pure tautology—the held-out image is unused for matching and the behavioral outcome can fail—but success is measured only for the fitted phenotype the optimizer targeted, so the synthesis 'prediction' is statistically favored rather than tested in unmatched samples. Abstract/strongest claim still present this as the synthesis result.

full rationale

The paper's derivation chain is largely non-circular. Population-specific ANN decoders are trained on Wang & Adolphs image-level judgments, then applied to a new stimulus set (MSFDE) and tested in an independent lab cohort; selected-versus-random separation is an empirical outcome, not a fit renamed as prediction. Closed-loop synthesis optimizes GAN latent codes under those same predictors and is then behaviorally retested. The only load-bearing soft spot is validation design: leave-one-image-out correlation matching reselects Prolific participants to the original group response templates before measuring gap reduction on held-out pairs. That scopes the claim to the image-level phenotype the models were trained to capture (as the Discussion acknowledges) and raises a mild fitted-subspace concern, but the held-out image is never used for matching and the measured |ASD–NT| change remains free empirical data—not equal to the training labels or the synthesis loss by construction. Self-citations (e.g., Kar 2022) supply related prior framing, not uniqueness theorems that force the present results. Methods/Results disagreement on the synthesis loss is a correctness inconsistency, not circularity. Overall score 3: one non-tautological but self-referential validation step; central selection claim is clean.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 3 invented entities

The central claims rest on standard vision-model transfer assumptions, group-averaged behavioral targets, a constrained generative emotion manifold, and a phenotype-matching validation rule. Free parameters are mostly decoder/optimization choices and selection thresholds rather than a single fitted constant that forces the result. Invented entities are methodological constructs (diagnostic image sets, population-specific readouts, gap-reducing synthesis), not new physical objects.

free parameters (5)
  • ridge-regression regularization strength for ASD and NT decoders
    Cross-validated ridge maps penultimate features to mean happy probability; the penalty and any implicit scaling are fit to Wang & Adolphs image-level means and determine predicted Δ.
  • feature layer choice (penultimate / visual embedding)
    Main selection uses penultimate layers after layer-wise MSFDE gap predictivity (Figure S2); this is a data-informed design choice that affects which images rank as diagnostic.
  • synthesis SGD halt criteria (ε=0.0001 or 25 iterations)
    Latent-code optimization stops by hand-set loss threshold or iteration cap, shaping how far faces move on the GANmut manifold.
  • phenotype-matching correlation template and leave-one-image-out selection rule
    Which online participants enter the synthesis test depends on correlation to original group templates; this selection rule is free relative to a fixed full-cohort analysis.
  • AU pixel-difference thresholding / 17-region masks from neutral references
    Descriptive AU summaries depend on AUCanvas masks and signed mean pixel changes; not load-bearing for the main behavioral claims but free for interpretability analyses.
axioms (6)
  • domain assumption Fixed pretrained vision backbones (AlexNet, VGG-19, ResNet-50, ConvNeXt, CORNet-S, ViT, CLIP) provide image embeddings that can be linearly decoded into population-specific emotion judgments.
    Methods: penultimate/visual embeddings + ridge readouts; standard transfer-learning assumption in computational vision/neuroscience.
  • domain assumption Group-mean image-level happy probabilities are adequate targets for discovering stimuli that separate autistic and neurotypical observers.
    Training uses mean ratings per group (Wang & Adolphs); Discussion Limitations notes individual-level models are future work.
  • domain assumption GANmut’s 2D polar latent space (θ emotion category, ρ intensity) can express behaviorally relevant facial transformations while preserving identity and realism.
    Image synthesis pipeline and Limitations: transformations are model-accessible, not exhaustive of all facial features.
  • domain assumption Self-reported formal autism diagnosis plus SRS/AQ on Prolific, after consistency checks, sufficiently identifies autistic vs neurotypical online groups for synthesis validation.
    Participants Methods online cohort; Limitations acknowledges self-report constraints.
  • ad hoc to paper Leave-one-image-out correlation matching to original group templates tests synthesis without circular use of the held-out image.
    Results Closed-loop synthesis / Figure 4C: primary validation metric is phenotype-matched, not full unmatched cohort.
  • standard math Standard ridge regression, SGD, and Pearson/Spearman statistics apply to these behavioral vectors.
    Used throughout decoder training, synthesis optimization, and reported correlations/tests.
invented entities (3)
  • Population-specific ANN behavioral decoders (ASD vs NT readouts on shared visual features) independent evidence
    purpose: Predict image-level happy judgments separately for each group and define a model-predicted diagnostic gap Δ.
    Methodological construct built from existing ANNs plus ridge; independent evidence is the prospective selection and synthesis behavioral tests, not a new biological object.
  • Diagnostic / high-leverage facial-expression stimuli (sparse image-level phenotype) independent evidence
    purpose: Name the small subset of faces that drive autistic–neurotypical separation and serve as selection/synthesis targets.
    Operationally defined by |ASD–NT| response differences and model rankings; supported by cross-validated sparsity plots (Figures 1B, 3E) but not a new latent trait measure with external clinical validation.
  • Gap-reducing closed-loop synthesized faces no independent evidence
    purpose: Transform diagnostic base images toward greater predicted group agreement for intervention-style validation.
    Generated inside GANmut under the population-specific objective; behavioral effect shown only under phenotype matching.

pith-pipeline@v1.1.0-grok45 · 26597 in / 3865 out tokens · 51775 ms · 2026-07-10T05:48:46.415524+00:00 · methodology

0 comments
read the original abstract

Understanding perceptual differences between autistic and neurotypical adults requires behavioral assays that are sensitive, reliable, and mechanistically informative. Facial emotion perception is a useful test case because group differences have been reported, but findings vary across studies. Here we show that this variability may reflect image-level sparsity: autistic-neurotypical differences in emotion judgments were concentrated in a small subset of diagnostic facial expressions rather than spread uniformly across stimuli. We trained population-specific artificial neural network models to predict image-level judgments for autistic and neurotypical participants, then used these models to select novel faces predicted to maximize group separation. In an independent cohort, model-selected images produced larger behavioral differences than matched random images. We then used the same models with a generative adversarial network to transform diagnostic images toward greater predicted group agreement. In phenotype-matched validation, synthesized images reduced behavioral separation relative to their matched originals. These results establish a model-guided framework for discovering and transforming stimuli that reveal population-specific perceptual differences. More broadly, they show how behavioral phenotyping can move beyond averaging across fixed stimulus sets toward optimized assays that identify the conditions under which neurodivergent perception diverges or converges.

Figures

Figures reproduced from arXiv: 2607.08533 by Kohitij Kar, Kushin Mukherjee, Maren Wehrheim, Na Yeon Kim, Ralph Adolphs.

Figure 1
Figure 1. Figure 1: Differences in facial emotion discrimination emerge from a small subset of diagnostic images. A. [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Artificial neural networks (ANNs) trained on behavioral data predict diagnostic images in new participants. A.Training ANN models on existing data. Behavioral responses and facial-image stimuli from Wang & Adolphs (2017) were used to train separate neurotypical (NT) and autistic (ASD) behavioral decoders for each ANN architecture. Feature activations from the penultimate layer of each network (AlexNet, VGG… view at source ↗
Figure 3
Figure 3. Figure 3: Behavior-aligned ANNs prospectively identify diagnostic facial-expression stimuli. A. Behavioral paradigm used to test ANN-selected and randomly selected images in an independent cohort of autistic and neurotypical adults. Each trial began with a 1000-ms fixation, followed by a brief test image presented for 100 ms at an 8° visual angle, and a two-alternative emotion-choice screen. Participants judged whet… view at source ↗
Figure 4
Figure 4. Figure 4: Closed-loop synthesis reduces autistic–neurotypical behavioral divergence. A. Closed-loop image￾synthesis pipeline for generating gap-reduced facial expressions. Starting from a diagnostic base image, GANmut was used to synthesize candidate facial expressions by updating the latent emotion code. Each candidate image was passed through the trained population-specific ANN behavioral models, which predicted “… view at source ↗
Figure 5
Figure 5. Figure 5: Gap-reducing synthesis produces structured facial-region transformations. A. Example of a gap￾reducing image transformation. The original diagnostic image is shown on the left, and the synthesized image is shown on the right. Colored overlays indicate pixels that changed relative to the original image after thresholding the absolute pixel-wise difference. Red and blue indicate positive and negative changes… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

68 extracted references · 68 canonical work pages · 9 internal anchors

  1. [1]

    Lord, C. et al. Autism spectrum disorder. Nat. Rev. Dis. Primer 6, 5 (2020)

    work page 2020

  2. [2]

    & Humphreys, K

    Behrmann, M., Thomas, C. & Humphreys, K. Seeing it differently: visual processing in autism. Trends Cogn. Sci. 10, 258–264 (2006)

    work page 2006

  3. [3]

    Robertson, C. E. & Baron-Cohen, S. Sensory perception in autism. Nat. Rev. Neurosci. 18, 671–684 (2017)

    work page 2017

  4. [4]

    & Yashar, A

    Hadad, B.-S. & Yashar, A. Sensory Perception in Autism: What Can We Learn? Annu. Rev. Vis. Sci. 8, 239–264 (2022)

    work page 2022

  5. [5]

    Zilbovicius, M. et al. Autism, the superior temporal sulcus and social perception. Trends Neurosci. 29, 359–366 (2006)

    work page 2006

  6. [6]

    Morrison, K. E. et al. Psychometric Evaluation of Social Cognitive Measures for Adults with Autism. Autism Res. 12, 766–778 (2019)

    work page 2019

  7. [7]

    Kennedy, D. P. & Adolphs, R. Perception of emotions from facial expressions in high- functioning adults with autism. Neuropsychologia 50, 3313–3319 (2012)

    work page 2012

  8. [8]

    Loth, E. et al. Facial expression recognition as a candidate marker for autism spectrum disorder: how frequent and severe are deficits? Mol. Autism 9, 7 (2018)

    work page 2018

  9. [9]

    Sato, W. et al. Impaired detection of happy facial expressions in autism. Sci. Rep. 7, 13340 (2017)

    work page 2017

  10. [10]

    & Adolphs, R

    Wang, S. & Adolphs, R. Reduced specificity in emotion judgment in people with autism spectrum disorder. Neuropsychologia 99, 286–295 (2017)

    work page 2017

  11. [11]

    M., Moody, E

    Beall, P. M., Moody, E. J., McIntosh, D. N., Hepburn, S. L. & Reed, C. L. Rapid facial reactions to emotional facial expressions in typically developing children and children with autism spectrum disorder. J. Exp. Child Psychol. 101, 206–223 (2008)

    work page 2008

  12. [12]

    Yi, L. et al. Abnormality in face scanning by children with autism spectrum disorder is limited to the eye region: Evidence from multi-method analyses of eye tracking data. J. Vis. 13, 5– 5 (2013). 27

    work page 2013

  13. [13]

    L., Piven, J

    Neumann, D., Spezio, M. L., Piven, J. & Adolphs, R. Looking you in the mouth: abnormal gaze in autism resulting from impaired top-down modulation of visual attention. Soc. Cogn. Affect. Neurosci. 1, 194–202 (2006)

    work page 2006

  14. [14]

    Dalton, K. M. et al. Gaze fixation and the neural circuitry of face processing in autism. Nat. Neurosci. 8, 519–526 (2005)

    work page 2005

  15. [15]

    A., Morris, J

    Pelphrey, K. A., Morris, J. P., McCarthy, G. & LaBar, K. S. Perception of dynamic changes in facial affect and identity in autism. Soc. Cogn. Affect. Neurosci. 2, 140–149 (2007)

    work page 2007

  16. [16]

    Kleinhans, N. M. et al. fMRI evidence of neural abnormalities in the subcortical face processing system in ASD. NeuroImage 54, 697–704 (2011)

    work page 2011

  17. [17]

    Rutishauser, U. et al. Single-Neuron Correlates of Atypical Face Processing in Autism. Neuron 80, 887–899 (2013)

    work page 2013

  18. [18]

    Griffin, J. W. et al. Decoding the temporal dynamics of face-specific neural representations in autism. Nat. Ment. Health 4, 1109–1121 (2026)

    work page 2026

  19. [19]

    & Hamilton, A

    Uljarevic, M. & Hamilton, A. Recognition of Emotions in Autism: A Formal Meta-Analysis. J. Autism Dev. Disord. 43, 1517–1526 (2013)

    work page 2013

  20. [20]

    López Pérez, D. et al. Visual Search Performance Does Not Relate to Autistic Traits in the General Population. J. Autism Dev. Disord. 49, 2624–2631 (2019)

    work page 2019

  21. [21]

    & Wagemans, J

    Evers, K., Steyaert, J., Noens, I. & Wagemans, J. Reduced Recognition of Dynamic Facial Emotional Expressions and Emotion-Specific Response Bias in Children with an Autism Spectrum Disorder. J. Autism Dev. Disord. 45, 1774–1784 (2015)

    work page 2015

  22. [22]

    & DiCarlo, J

    Kar, K. & DiCarlo, J. J. The Quest for an Integrated Set of Neural Mechanisms Underlying Object Recognition in Primates. Annu. Rev. Vis. Sci. 10, 91–121 (2024)

    work page 2024

  23. [23]

    Schrimpf, M. et al. Brain-Score: Which Artificial Neural Network for Object Recognition Is Most Brain-Like? http://biorxiv.org/lookup/doi/10.1101/407007 (2018) doi:10.1101/407007

    work page doi:10.1101/407007 2018

  24. [24]

    Rajalingham, R. et al. Large-Scale, High-Resolution Comparison of the Core Visual Object Recognition Behavior of Humans, Monkeys, and State-of-the-Art Deep Artificial Neural Networks. J. Neurosci. 38, 7255–7269 (2018)

    work page 2018

  25. [25]

    Yamins, D. L. K. et al. Performance-optimized hierarchical models predict neural responses in higher visual cortex. Proc. Natl. Acad. Sci. 111, 8619–8624 (2014). 28

    work page 2014

  26. [26]

    & DiCarlo, J

    Bashivan, P., Kar, K. & DiCarlo, J. J. Neural population control via deep image synthesis. Science 364, eaav9436 (2019)

    work page 2019

  27. [27]

    & Tsao, D

    Chang, L. & Tsao, D. Y. The Code for Facial Identity in the Primate Brain. Cell 169, 1013- 1028.e14 (2017)

    work page 2017

  28. [28]

    Snoek, L. et al. Testing, explaining, and exploring models of facial expressions of emotions. Sci. Adv. 9, eabq8421 (2023)

    work page 2023

  29. [29]

    T., Ramezanpour, H

    Wehrheim, M., Alamooti, S. T., Ramezanpour, H. & Kar, K. Facial expression discrimination emerges from neural subspaces shared with detection and identity. Preprint at https://doi.org/10.1101/2025.08.25.672186 (2025)

    work page doi:10.1101/2025.08.25.672186 2025

  30. [30]

    A Computational Probe into the Behavioral and Neural Markers of Atypical Facial Emotion Processing in Autism

    Kar, K. A Computational Probe into the Behavioral and Neural Markers of Atypical Facial Emotion Processing in Autism. J. Neurosci. 42, 5115–5126 (2022)

    work page 2022

  31. [31]

    Yu, H. et al. Multimodal investigations of emotional face processing and social trait judgment of faces. Ann. N. Y. Acad. Sci. 1531, 29–48 (2024)

    work page 2024

  32. [32]

    A., Bashivan, P., Abate, A., DiCarlo, J

    Ratan Murty, N. A., Bashivan, P., Abate, A., DiCarlo, J. J. & Kanwisher, N. Computational models of category-selective brain regions enable high-throughput tests of selectivity. Nat. Commun. 12, 5540 (2021)

    work page 2021

  33. [33]

    Ponce, C. R. et al. Evolving Images for Visual Neurons Using a Deep Generative Network Reveals Coding Principles and Neuronal Preferences. Cell 177, 999-1009.e10 (2019)

    work page 2019

  34. [34]

    P., Huang, Z., Romero, A

    d’Apolito, S., Paudel, D. P., Huang, Z., Romero, A. & Gool, L. V. GANmut: Learning Interpretable Conditional Space for Gamut of Emotions. in 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 568–577 (IEEE, Nashville, TN, USA, 2021). doi:10.1109/CVPR46437.2021.00063

    work page doi:10.1109/cvpr46437.2021.00063 2021

  35. [35]

    Tuckute, G. et al. How to optimize neuroscience data utilization and experiment design for advancing brain models of visual and linguistic cognition? Preprint at https://doi.org/10.48550/ARXIV.2401.03376 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2401.03376 2024

  36. [36]

    & Hinton, G

    Krizhevsky, A., Sutskever, I. & Hinton, G. E. ImageNet Classification with Deep Convolutional Neural Networks. in Advances in Neural Information Processing Systems (eds Pereira, F., Burges, C. J. C., Bottou, L. & Weinberger, K. Q.) vol. 25 (Curran Associates, Inc., 2012). 29

    work page 2012

  37. [37]

    Very Deep Convolutional Networks for Large-Scale Image Recognition

    Simonyan, K. & Zisserman, A. Very Deep Convolutional Networks for Large-Scale Image Recognition. ArXiv14091556 Cs http://arxiv.org/abs/1409.1556 (2015)

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  38. [38]

    Deep Residual Learning for Image Recognition

    He, K., Zhang, X., Ren, S. & Sun, J. Deep Residual Learning for Image Recognition. Preprint at https://doi.org/10.48550/arXiv.1512.03385 (2015)

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1512.03385 2015

  39. [39]

    Woo, S. et al. ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders. Preprint at https://doi.org/10.48550/ARXIV.2301.00808 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2301.00808 2023

  40. [40]

    Kubilius, J. et al. Brain-Like Object Recognition with High-Performing Shallow Recurrent ANNs. in Advances in Neural Information Processing Systems (eds Wallach, H. et al.) vol. 32 (Curran Associates, Inc., 2019)

    work page 2019

  41. [41]

    Dosovitskiy, A. et al. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. Preprint at https://doi.org/10.48550/arXiv.2010.11929 (2021)

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2010.11929 2010

  42. [42]

    Radford, A. et al. Learning Transferable Visual Models From Natural Language Supervision. Preprint at https://doi.org/10.48550/arXiv.2103.00020 (2021)

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2103.00020 2021

  43. [43]

    Beaupré, M. G. & Hess, U. Cross-Cultural Emotion Recognition among Canadian Ethnic Groups. J. Cross-Cult. Psychol. 36, 355–370 (2005)

    work page 2005

  44. [44]

    Doerig, A. et al. High-level visual representations in the human brain are aligned with large language models. Nat. Mach. Intell. 7, 1220–1234 (2025)

    work page 2025

  45. [45]

    Muttenthaler, L., Dippel, J., Linhardt, L., Vandermeulen, R. A. & Kornblith, S. Human alignment of neural network representations. Preprint at https://doi.org/10.48550/arXiv.2211.01201 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2211.01201 2025

  46. [46]

    Kar, K., Kubilius, J., Schmidt, K., Issa, E. B. & DiCarlo, J. J. Evidence that recurrent circuits are critical to the ventral stream’s execution of core object recognition behavior. Nat. Neurosci. 22, 974–983 (2019)

    work page 2019

  47. [47]

    & Kar, K

    Muzellec, S. & Kar, K. Reverse predictivity for bidirectional comparison of neural networks and biological brains. Nat. Mach. Intell. 8, 474–488 (2026)

    work page 2026

  48. [48]

    Sörensen, L. K. A., DiCarlo, J. J. & Kar, K. Hierarchical optimization predicts plasticity in the macaque inferior temporal cortex following object training. Nat. Commun. https://doi.org/10.1038/s41467-026-74816-0 (2026) doi:10.1038/s41467-026-74816-0. 30

    work page doi:10.1038/s41467-026-74816-0 2026

  49. [49]

    & Gosselin, F

    Blais, C., Roy, C., Fiset, D., Arguin, M. & Gosselin, F. The eyes are not the window to basic emotions. Neuropsychologia 50, 2830–2838 (2012)

    work page 2012

  50. [50]

    & Plomin, R

    Happé, F., Ronald, A. & Plomin, R. Time to give up on a single explanation for autism. Nat. Neurosci. 9, 1218–1220 (2006)

    work page 2006

  51. [51]

    Geschwind, D. H. Advances in Autism. Annu. Rev. Med. 60, 367–380 (2009)

    work page 2009

  52. [52]

    Lai, M.-C., Lombardo, M. V. & Baron-Cohen, S. Autism. The Lancet 383, 896–910 (2014)

    work page 2014

  53. [53]

    & Breakspear, M

    Opel, N. & Breakspear, M. Transforming mental health research and care through artificial intelligence. Science 391, 249–258 (2026)

    work page 2026

  54. [54]

    K., Kornblith, S

    Muzellec, S., Alghetaa, Y. K., Kornblith, S. & Kar, K. MAPS: Masked Attribution-based Probing of Strategies- A computational framework to align human and model explanations. Preprint at https://doi.org/10.48550/arXiv.2510.12141 (2025)

    work page doi:10.48550/arxiv.2510.12141 2025

  55. [55]

    GANalyze: Toward Visual Definitions of Cognitive Image Properties

    Goetschalckx, L., Andonian, A., Oliva, A. & Isola, P. GANalyze: Toward Visual Definitions of Cognitive Image Properties. ArXiv190610112 Cs http://arxiv.org/abs/1906.10112 (2019)

    work page internal anchor Pith review Pith/arXiv arXiv 1906

  56. [56]

    Controlling Memorability of Face Images

    Younesi, M. & Mohsenzadeh, Y. Controlling Memorability of Face Images. Preprint at http://arxiv.org/abs/2202.11896 (2022)

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  57. [57]

    G., Skora, L

    Krumhuber, E. G., Skora, L. I., Hill, H. C. H. & Lander, K. The role of facial movements in emotion recognition. Nat. Rev. Psychol. 2, 283–296 (2023)

    work page 2023

  58. [58]

    & Todorov, A

    Aviezer, H., Trope, Y. & Todorov, A. Body Cues, Not Facial Expressions, Discriminate Between Intense Positive and Negative Emotions. Science 338, 1225–1229 (2012)

    work page 2012

  59. [59]

    & Gosselin, F

    Belin, P., Fillion-Bilodeau, S. & Gosselin, F. The Montreal Affective Voices: A validated set of nonverbal affect bursts for research on auditory affective processing. Behav. Res. Methods 40, 531–539 (2008)

    work page 2008

  60. [60]

    & Lord, C

    Hus, V. & Lord, C. The Autism Diagnostic Observation Schedule, Module 4: Revised Algorithm and Standardized Severity Scores. J. Autism Dev. Disord. 44, 1996–2012 (2014)

    work page 1996

  61. [61]

    N., Penton-Voak, I

    Dalili, M. N., Penton-Voak, I. S., Harmer, C. J. & Munafò, M. R. Meta-analysis of emotion recognition deficits in major depressive disorder. Psychol. Med. 45, 1135–1144 (2015)

    work page 2015

  62. [62]

    & Schitter, C

    Palan, S. & Schitter, C. Prolific.ac—A subject pool for online experiments. J. Behav. Exp. Finance 17, 22–27 (2018). 31

    work page 2018

  63. [63]

    & Shapiro, D

    Chandler, J. & Shapiro, D. Conducting Clinical Research Using Crowdsourced Convenience Samples. Annu. Rev. Clin. Psychol. 12, 53–81 (2016)

    work page 2016

  64. [64]

    Constantino, J. N. Social Responsiveness Scale. in Encyclopedia of Autism Spectrum Disorders (ed. Volkmar, F. R.) 2919–2929 (Springer New York, New York, NY, 2013). doi:10.1007/978-1-4419-1698-3_296

    work page doi:10.1007/978-1-4419-1698-3_296 2013

  65. [65]

    & Clubley, E

    Baron-Cohen, S., Wheelwright, S., Skinner, R., Martin, J. & Clubley, E. The Autism- Spectrum Quotient (AQ): Evidence from Asperger Syndrome/High-Functioning Autism, Males and Females, Scientists and Mathematicians. J. Autism Dev. Disord. 31, 5–17 (2001)

    work page 2001

  66. [66]

    Choi, Y. et al. StarGAN: Unified Generative Adversarial Networks for Multi-domain Image- to-Image Translation. in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition 8789–8797 (IEEE, Salt Lake City, UT, 2018). doi:10.1109/CVPR.2018.00916

    work page doi:10.1109/cvpr.2018.00916 2018

  67. [67]

    M., Sanfeliu, A

    Pumarola, A., Agudo, A., Martinez, A. M., Sanfeliu, A. & Moreno-Noguer, F. GANimation: Anatomically-Aware Facial Animation from a Single Image. in Computer Vision – ECCV 2018 (eds Ferrari, V., Hebert, M., Sminchisescu, C. & Weiss, Y.) vol. 11214 835–851 (Springer International Publishing, Cham, 2018)

    work page 2018

  68. [68]

    & Gunes, H

    Luo, C., Song, S., Xie, W., Shen, L. & Gunes, H. Learning Multi-dimensional Edge Feature- based AU Relation Graph for Facial Action Unit Recognition. in Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence 1239–1246 (International Joint Conferences on Artificial Intelligence Organization, Vienna, Austria, 2022). doi:10...

    work page doi:10.24963/ijcai.2022/173 2022