The Computational Role of the Cortical multi-area Structure – CRoCoS

Mathematical models of neural networks are a key tool for understanding how millions of neurons in the brain work together to implement behaviorally relevant computations. Recent progress in artificial intelligence has provided new methods for generating network models that perform the same cognitive tasks as investigated with animals in systems neuroscience. Analyzing and interpreting such model networks has led to important insights into how computations emerge from collective dynamics of neural activity, yet the available models often lack key aspects of the biological structure of the brain.

A prominent level of structure in the mammalian cortex is the organization into anatomically-defined regions, and a premise of much of cognitive neuroscience is that different regions play different computational roles. A mechanistic understanding of the functional significance of the multi-area structure however remains elusive due to the paucity of multi-area network models and multi-area electrophysiological recordings to constrain them.

In this project we will develop a new class of mathematically tractable multi-area recurrent neural networks that we will compare with recordings in behaving ferrets to identify the constraints that complex behavioral tasks impose on the structure of activity. The central goal of the project will be to test the hypothesis that a multi-area structure endows a network with compositionality, i.e. the capacity to robustly perform multiple tasks by combining simpler modularized computations.

To this end, we will combine three complementary methods: (i) mathematical analyses of network dynamics;(ii) reverse-engineering of networks trained using back-propagation to perform multiple tasks; (iii) comparisons of predictions generated by model networks with cortical activity recorded in animals performing identical tasks.