Feature Visualization Recovers Brain Selectivity from TRIBE v2

Brain encoder models predict cortical fMRI responses from pretrained vision networks but are typically evaluated on prediction accuracy alone—a poor measure of interpretability. This paper, authored by Stuart Bladon and Brinnae Bent, introduces feature visualization as a complementary technique: gradient ascent on the encoder's predicted activation for a target brain region (ROI). Using the TRIBE v2 encoder paired with V-JEPA 2 (40-layer ViT-G), they synthesize still images for seven visual regions (V1 through ventral/dorsal areas).

Under identical hyperparameters, the probe recovers a clear progression of spatial scale and complexity across V1 to V4, matching the ventral-stream hierarchy. It also yields three distinctive downstream regimes: radial "frozen-motion" streaks for MT (despite static-only optimization), face-like features for FFA, and rectilinear line patterns for PPA. Notably, optimized FFA stimuli drive predicted region activation ~4x that of a natural face photograph, indicating these are adversarial super-stimuli rather than canonical exemplars. The method is simple, differentiable, and broadly applicable to any brain encoder with a differentiable backbone.