Beyond Prediction Accuracy: Target-Space Recovery Profiles for Evaluating Model-Brain Alignment
Researchers have developed a new framework to evaluate how well artificial vision models align with the human visual cortex. This method goes beyond simple prediction accuracy to analyze which specific dimensions of brain responses are recovered by the models. By using fMRI data from subjects viewing images, the study identified reproducible response dimensions in the visual cortex and assessed how effectively models and other brains recovered these dimensions. The findings suggest that prediction accuracy alone can obscure mismatches, and this new approach offers a more diagnostic evaluation of model-brain alignment. AI
IMPACT Provides a more nuanced evaluation of AI vision models' understanding of human visual processing.