Figure 1.

Trained RNN to perform the delayed context-dependent integration task

(A) Behavioral task description. A monkey was trained to discriminate, depending on the contextual cue, either the predominant color or predominant motion direction of randomly moving dots, and to further indicate its decision with a saccadic eye movement to a choice target. The cue stimulus onset determined the current context, which was characterized by different shapes and colors of the fixation point. The cue stimulus was followed by a fixed-delay epoch, and then followed by randomly moving dots stimuli. The monkey was rewarded for a saccade to the target matching the current context.

(B) Schematic of a fully connected, nonlinear RNN in context-dependent computation. The network received noisy inputs of two types: time-varying sensory stimulus and cue input. The stimulus inputs consisted of four task-relevant sensory information, each represented by an input unit that encoded the evidence for the direction of stimulus. The cue inputs consisted of two cue signals, which were represented by two input units (that indicated the current context and instructed the network to distinguish the type of stimulus). The two output channels encoded the response direction.

(C and D) RNN learning curve (C) and the performance curve (D). Training was completed once both quantities reached the convergence criterion (blue horizontal dashed lines).

(E) Psychometric curves in a delayed context-dependent integration task. The probability of a correct direction judgment is plotted as a function of color (Left) and motion (Right) coherence in color-context (blue) and motion-context (black) trials.

(F) The activities of representative units indicated by different colors. The first gray shading area indicates cue stimulus epoch and the second shading area indicates the presentation of random dots (i.e., integration of sensory stimulus epoch)