Spatiotemporal Encoding of Task Variables Across the Mouse Brain
brain-wide decoding; dimensionality reduction; latent space; decision-making.
This thesis investigates the spatiotemporal organization of neural representations underlying decision-making across the mouse brain. Leveraging a large-scale electrophysiological dataset from the International Brain Laboratory (IBL), encompassing recordings from 115 mice and 267 brain regions, we employed logistic regression to decode behavioral variables during a visual decision-making task. Our brain-wide decoding analysis revealed a nuanced interplay of localization and distribution: choice and feedback were broadly decodable across numerous regions, indicating distributed encoding, while visual stimulus and contrast exhibited more localized decodability, suggesting specialized sensory processing. Furthermore, temporal dynamics analysis revealed distinct profiles for each variable: choice decoding peaked perimovement, feedback showed sustained post-feedback decodability, and visual stimuli displayed delayed, gradual encoding. Complementing direct decoding, an embeddings consistency analysis using CEBRA for dimensionality reduction confirmed the robustness of these latent representations. These findings provide a comprehensive brain-wide map of decision-making neural correlates, elucidating the complex spatiotemporal dynamics and distributed versus localized nature of neural representations underlying distinct behavioral variables.