Fig. 4. Architecture of the RNN models.

a The models consist of three layers with different types of units mimicking different populations of neurons: sensory units, recurrent units, and an output unit. The recurrent units (N = 120) included both excitatory and inhibitory populations that receive input from 63 sensory populations encoding individual features (N = 9), conjunctions of features (N = 27), and object-identity of each stimulus (N = 27). Among the recurrent populations only the excitatory recurrent populations project to the output population. Half of the connections from sensory populations and the connections between recurrent populations were endowed with reward-dependent plasticity. b Based on the type of populations and uniform presence/absence of reward-dependent plasticity in the connections to and between recurrent populations, these populations could be grouped into eight disjoint populations: Exc_rr and Inh_rr corresponding to populations with no plastic sensory or recurrent connections (rigid weights indicated by subscript r); Exc_fr and Inh_fr corresponding to populations with plastic sensory input only (flexible weights indicated by subscript f); Exc_rf and Inh_rf corresponding to populations with plastic recurrent connections only; and Exc_ff and Inh_ff corresponding to populations with plastic sensory input and plastic recurrent connections. c Activity of the sensory and recurrent populations in an example trial. Upon presentation of a stimulus, three feature-encoding, three conjunction-encoding, and one object-identity encoding populations become active and give rise to activity in the excitatory and inhibitory recurrent populations. Source data are provided as a Source Data file.