Unstructured representations in a structured brain: a cross-region analysis of the neural code

A major question in computational neuroscience is what computational strategies the brain employs to represent the external world, the internal states, and its decisions. An open debate is whether neurons are organized in specialized sub-populations to encode behaviorally relevant variables ("categorical" selectivity) or whether information on several variables is distributed across the neural population ("heterogeneous mixed" selectivity). With evidence being published in support of both views, it is unclear under which conditions the brain uses one or the other encoding strategy.

We will present an analysis approach to address these questions on neural activity recorded from ~30 cognitive and sensory cortical areas during a decision-making task. We will show that the encoding strategy depends on the scale on which neural populations are considered: on a larger inter-region scale, neurons are categorical, i.e., specialized into subpopulations for different combinations of variables. However, within individual regions, neurons are mostly mixed-selective, with few exceptions in somatosensory and visual areas.