A cortical information bottleneck during decision-making.

The brain uses multiple areas for cognition, decision-making, and action, but it is unclear why the brain distributes the computation and why cortical activity differs by brain area. Machine learning and information theory suggests that one benefit of multiple areas is that it provides an "information bottleneck" that compresses inputs into an optimal representation that is minimal and sufficient to solve the task. Combining experimental recordings from behaving animals and computational simulations, we show that later brain areas have a tendency to form such minimal sufficient representations of task inputs through preferential propagation of task-relevant information present in earlier areas. Our results thus provide insight into why the brain uses multiple brain areas for supporting decision-making and action.