Navve Wasserman

Navve will present two recent projects that develop automated and scalable approaches for discovering and interpreting visual representations in the human brain. BrainExplore (https://navvewas.github.io/BrainExplore/) addresses the challenge of searching the vast space of possible visual representations by using data-driven decomposition of fMRI activity to discover candidate neural patterns, identifying the natural images that most strongly activate them, and automatically generating and quantitatively evaluating natural-language explanations of their visual meaning. This framework enables large-scale discovery of thousands of interpretable patterns across the visual cortex. Building on the challenge of reliably identifying what these neural patterns actually represent, BrainCause (https://yuvalgol123.github.io/BrainCause/) moves from activation-based discovery toward causal validation: it uses generative models to construct controlled stimuli, including counterfactual edits and semantic negatives, to test whether a brain region responds specifically to a target visual concept rather than to correlated visual or semantic features.