Figure 7. Network models of ALM interhemispheric interactions.

(A) Recurrent neural network (RNN) consists of two modules (‘hemispheres’). Each module receives noisy sensory input and a linear readout unit generates choice activity.

(B) Activity of a robust modular RNN on “lick right” (blue) and “lick left” (red) trials. Solid lines, photoinhibition trials; dotted lines, control trials; shades, standard deviation across trials. Cyan bars, photoinhibition period.

(C) State-dependent gating in robust RNNs. Left, the strength of influence from the contra hemisphere as a function of choice encoding strength in both hemispheres (Methods). The influence is primarily modulated by choice encoding strength in the analyzed hemisphere. Red and blue dotted lines, average CD projection of correct “lick left” and “lick right” trials in the analyzed hemisphere; black dotted line, decision boundary. Right, the strength of interhemispheric influence when the analyzed hemisphere is in the weakly (cross) versus highly selective state (circle). 30 randomly initialized RNNs.

(D) Top, a modular RNN receiving sensory input of equal strength in each hemisphere. The plots show RNN activity during contralateral photoinhibition (c.f. Figure 2B). Bottom, an asymmetric RNN receiving stronger sensory input in one hemisphere.

(E) Modularity of each hemisphere in RNNs as a function of input asymmetry across hemispheres (1 or −1 indicates input only to the right or left hemisphere respectively). Error bar, standard error across RNNs (n = 50).

See also Figure S7.