Estimating flexible across-area communication with neurally-constrained RNN

Joao Barbosa, Adrian Valente, Scott L Brincat, Earl K Miller, Srdjan Ostojic

Abstract:

Neural computations supporting complex behaviors involve multiple brain regions, and large-scale recordings
from animals engaged in complex tasks are increasingly
common. A current challenge in analysing these data
is to identify which part of the information contained
within a brain region is shared with others. Here, to
address this limitation, we trained multi-region recurrent neural networks (RNN) models to reproduce the dynamics of large-scale single-unit recordings (more than
6000 neurons across 7 cortical areas) from monkeys engaged in a two-dimensional (color and motion direction)
context-dependent decision-making task. Decoding analyses show that all areas encode both stimuli (color and
direction). However, using our approach we uncovered
feed-forward and feedback interactions within a network
of 7 interacting regions. Constraining interactions during training or testing recovered the canonical brain hierarchy that differentiate sensory and frontal regions. Inspecting across-region interactions, we also found that
frontal regions compress the irrelevant stimulus in a
context-dependent manner, while sensory regions always
compress the same stimulus.