Fig. 1. Architecture of the network. The module consists in N neurons and is divided in two major groups: one inhibitory and one excitatory population, with N_I and N_E neurons each (N=N_E+N_I). Within the excitatory pool there are two types of populations: two selective pools A and B, each constituted by fN_E neurons (f=0.15), and one non-selective population, formed by all excitatory neurons not belonging to a selective pool ((1-2f)N_E). The firing rate activity of the two selective pools encode the decision to make. w₊ are the synaptic weights connecting neurons within the same selective pool, whereas w_- denotes the connection weight between neurons in different selective pools and from non-selective to selective neurons. All other possible connections have weight 1 (baseline strength). To assure that the overall recurrent excitatory synaptic drive in the spontaneous state remains constant as w₊ is modified, w_- is set to 1-f(w₊-1)/(1-f). The signals associated to the stimuli are denoted by λ and the signal carrying value information is (one per pool). Every neuron in the network receives a background Poisson spike train of rate .

Fig. 2. Mean-field exploration of the network's regimes of operation for w₊=1.7 and w₊=1.8. Each region corresponds to a regime of operation, determined by the set of stable states that the network can sustain. Stable states are denoted by their initials: S (spontaneous), M (mixed), A (A-selective), and B (B-selective). The shaded area represents the unphysical region, defined by . In the competition region (A,B) there are two stable selective states, and the network must take a binary decision. Note that the diagrams show only the region Δλ>0; the diagrams for Δλ<0 are the mirror images about Δλ=0 of the diagrams shown here, with A and B interchanged.

Fig. 3. (a) Difference between the asymptotic firing rates associated with a high and a low expected gain, Δν, with respect to the bias in the value signal, Δλ, for different values of the recurrent potentiation weight w₊ and with . The curves are plotted in the Δλ interval where there is competition. (b) Proportion of trials (out of 500) in which the network chose A as a function of the value signal received at A (), keeping the other value signal () fixed at 0. Negative values for should be interpreted as the symmetric situation, where only population B receives a value signal, equal to .