Assessing psychological, neural and AI-based representational geometries across multiple similarity properties
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Assessing similarities between objects is a fundamental cognitive process for humans. However, different similarity properties (e.g., color, animacy) elicit distinct behavioral judgments. In this study, we explored the relationships between similarity judgments (for 280 objects), fMRI-RSA matrices (11 brain regions), and 14 AI-model matrices trained on object-based properties across diverse similarity metrics. Our results revealed a behavioral distinction between “object-based” (e.g., shape) and “subject-based” (e.g., preference) representational geometries. Object-based categories significantly correlated with neural spaces, particularly from occipital and fusiform regions. However, these results were influenced by the consistency of activation patterns across brain regions, suggesting a higher granularity in sensory cortices compared to multimodal areas. Lastly, only DNNs with high classification performance on specific properties correlated significantly with behavioral and neural data. Our findings provide insights into different human representational spaces, their relationship with brain activity across regions, and under which conditions AI-models can resemble both.