01 Aligning foundation models and the brain
Foundation models learn rich representations and provide a useful basis for in silico experiments. Nevertheless, due to considerable differences in representation format (e.g., dense spatial features vs a handful of neurons) and available data (e.g., 2 hours of electrophysiological recordings), applying these models to interpret neural dynamics remains a major challenge. My research bridges this gap by developing self-supervised learning paradigm that anchors the two representations into a shared coordinate, preserving the structure developed in the foundation models while captureing the representational geometry in the neural activity.