Schedule of Events | Search Abstracts | Invited Symposia | Symposia | Poster Sessions | Data Blitz

Sketchpad Series 

VisDeep: Distribution-Matched Stimulus Selection Using Earth Mover’s Distance

Poster Session B - Sunday, March 8, 2026, 8:00 – 10:00 am PDT, Fairview/Kitsilano Ballroom

Deepkhushi Baidwan¹ (deepkhushibaidwan@gmail.com), Eric Mah, Jim Tanaka; ¹University of Victoria, ²Different Minds Lab

Selecting perceptual stimuli that are appropriately matched across experimental conditions remains a persistent challenge in cognitive research. Traditional approaches rely on subjective judgment or matching on a limited set of attributes, increasing the risk of hidden confounds and reduced reproducibility. This work introduces VisDeep, a Python-based framework designed to support objective, distribution-matched stimulus selection using high-dimensional perceptual representations. VisDeep integrates multimodal stimulus features, including physical measurements, subjective perceptual ratings, demographic attributes, and deep neural network embeddings, into a unified feature space. The core methodological contribution is an Earth Mover’s Distance (EMD)-based optimization procedure that searches across candidate stimulus subsets to identify groups with closely matched feature distributions while preserving within-group heterogeneity. The pipeline includes automated preprocessing, configurable feature weighting, and statistical evaluation tools to support transparent validation. As preliminary validation, VisDeep was applied to the Chicago Face Database, producing a complete dataset of 1,207 face stimuli represented by 4,172 features each. Using the EMD-based selection procedure, two demographically balanced face subsets (20 Asian and 20 White faces) were identified with closely matched distributions across neural, physical, and perceptual features (EMD = 0.032). These pilot results demonstrate the feasibility and stability of the proposed approach. Ongoing work uses VisDeep to compare machine-derived perceptual spaces with human psychological similarity spaces derived from behavioral judgments. Potential outcomes including partial alignment or systematic divergence will be interpreted in relation to known properties of human face perception. Although demonstrated with faces,VisDeep is designed to generalize to other perceptual domains, supporting broader methodological adoption.

Topic Area: METHODS: Other