AI-guided Stimuli Discovery and Generation to Optimize Facial Emotion Perception Studies in Autism

Jul 16, 2026

Autistic and neurotypical perceptual differences in facial emotion perception remain difficult to measure with assays that are both sensitive and mechanistic. Facial emotion tasks show inconsistent behavioral effects across tasks, individuals, emotions, and stimuli, suggesting the phenotype varies by stimulus. This work demonstrates that differences concentrate in a small subset of diagnostic images. Population-specific ANN models predict image-level judgments and prospectively screen stimuli, improving assay sensitivity. A constrained GAN then transforms diagnostic images to shift responses and test whether separation increases or decreases.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages37

Reading Time1 h 33 min

Abstract

Introduction

Title AI-guided stimuli discovery and generation to optimize facial emotion perception studies in autism Authors Kushin Mukherjee 1,2 *, Na Yeon Kim3 *, Maren Wehrheim 1,4, Ralph Adolphs5, and Kohitij Kar 1,4 Affiliation 1. Centre for Integrative and Applied Neuroscience, York University, Canada 2. Department of Psychology, Stanford University, USA 3. Department of Psychology, University of California, Riverside, USA 4. Department of Biology, and Centre for Vision Research, Toronto, York University, Canada 5. Division of the Humanities and Social Sciences, California Institute of Technology, Pasadena, CA, USA * contributed equally to this work. Correspondence should be addressed to Kohitij Kar [E-mail: k0h1t1j@yorku.ca](E-mail: k0h1t1j@yorku.ca) Conflict of interests The author declares no competing financial interests. Acknowledgments KK has been supported by funds from the Canada Research Chair Program, the Simons Foundation Autism Research Initiative (SFARI, 967073), Brain-Canada Foundation (2023- 0259), the Canada First Research Excellence Funds (VISTA Program), the National Sciences and Engineering Research Council of Canada (NSERC, RGPIN-2024-06223), and the Ontario Early Researcher Awards program (ER24-18-272) . NYK was supported by the Della Martin Foundation Postdoctoral Fellowship. RA was supported by the Simons Foundation Autism Research Initiative (SFARI, 990500) . Data and Code Availability Curated de-identified data and code for reproducing the reported figure panels will be available at [https://github.com/vital-kolab/autism-ann-emotion](https://github.com/vital-kolab/autism-ann-emotion upon publication)[ upon publication](https://github.com/vital-kolab/autism-ann-emotion upon publication). Abstract Understanding how perceptual differences arise between autistic and neurotypical individuals remains a major challenge in autism research. A key obstacle is the lack of behavioral assays that are both sensitive to these differences and sufficiently mechanistic to reveal the computations that underlie them. Facial emotion perception is a compelling domain for making progress on assay development because differences between autistic and neurotypical observers have been widely reported, yet the mechanisms that produce these differences remain unresolved. Some studies find robust group differences in emotion judgments, whereas others report weak or inconsistent effects, suggesting that the behavioral phenotype is not expressed uniformly across tasks, individuals, or stimuli. Here, we show that autistic–neurotypical differences in facial emotion perception are not uniformly distributed across images, but are concentrated in a small subset of diagnostic stimuli. We trained population-specific artificial neural network (ANN) models to predict image-level emotion judgments in autistic and neurotypical participants and used these models to prospectively screen novel facial expressions. ANN-guided image selection identified stimuli that produced larger autistic– neurotypical behavioral separation than randomly selected images in an independent cohort, demonstrating that computational models can improve assay sensitivity by selecting highleverage stimuli. We then extended this framework from stimulus selection to stimulus transformation. Using a generative adversarial network constrained by the ANN-based behavioral models, we synthesized modified versions of diagnostic images designed to move responses toward greater autistic–neurotypical agreement. In a phenotype-matched validation analysis, these synthesized images reduced behavioral separation relative to their matched base images. Together, these findings establish a model-guided framework for discovering and transforming stimuli that reveal population-specific perceptual differences. More broadly, this approach shifts behavioral phenotyping from averaging across fixed stimulus sets toward computationally optimized assays that identify

FAQs

Why are facial emotion perception assays challenging in autism research?

Behavioral assays often lack both sensitivity to autistic–neurotypical differences and mechanistic specificity. Reported effects also vary strongly across tasks, individuals, emotions, and stimuli.

How does the study identify which facial expressions are most informative?

It trains population-specific ANN models to predict image-level emotion judgments for autistic and neurotypical participants, then uses these models to prospectively screen novel facial expressions for higher “behavioral separation.”

What does the framework do after stimulus selection?

It extends from selecting images to transforming them using a constrained generative adversarial network. Synthesized image versions are designed to move responses toward greater autistic–neurotypical agreement, and validation shows reduced separation versus matched base images.

Jul 16, 2026

Upload by Valentina