Our models predict that in different regimes circuits have divergent responses to perturbations of inhibitory neurons: when competitive inhibition dominates suppressing inhibition will slow decisions, whereas decisions will be faster when stabilizing inhibition dominates.
In RNNs trained to perform a decision-making task, connections between choice selective units had the same relationship between the specificity of recurrent excitation and whether selective inhibition was contraspecific or ipsispecific.
The two alternative roles for inhibition—competitive or stabilizing—are apparent in the performance of the circuits. Ipsispecific inhibition leads to slow, accurate decisions while contraspecific inhibition leads to faster decisions at the expense of accuracy.
The specificity of inhibitory connections can compensate for the specificity of excitation: when inhibition is contraspecific, weakly specific excitatory circuits can make decisions and when inhibition is ipsispecific, strongly specific excitatory circuits support decisions.
We constructed circuit models with excitatory and inhibitory populations selective for one of two choices. We connect inhibitory outputs in either contraspecific (across choice) or ipsispecific (within choice) motif.
