Fig. 2: Ring attractor network with synaptic STP shows serial dependence.

Simulations of two consecutive working memory trials (current trial n, previous trial n − 1) in a spiking neural network model with bump-attractor dynamics (“Methods”). a Spike times (x-axis) of excitatory neurons, ordered on y-axis by preferred angular location. Colored bars in a, b mark previous and current stimulus onset times (olive) and previous response (red). The solid orange line shows the population vector decoded from firing rates (sliding windows of 250 ms). In trial n, the active memory representation got biased towards the memory representation in trial n − 1. b Firing rate (black) and potentiated weight trace w_ij for neurons at 0° (orange) averaged over 1,000 trials and 20 neurons centered around 0°. Spiking activity and synaptic strength increased during trial n − 1 delay and decreased after the response. At current stimulus onset, information about trial n − 1 remained only in the potentiated weight trace. To facilitate interpretation, we excluded trials for which any neuron participated in previous and current-trial delay activity (i.e., showed firing rates >10 spikes s⁻¹ after stimulus onset in trial n). c, d Associativity and decay of modeled STP. The strength of each individual synapse is determined by w_ij (c, middle black trace), which is potentiated at each spike by an amount Δ_w that depends on the relative spike times t_j and t_i of pre- and postsynaptic neurons, respectively, and on the potentiation factor P that is chosen to represent different strengths of STP (different colored lines in (d); “Methods”, Eqs. (15) and (16)), and it is reduced by an amount relative to the synaptic strength at each presynaptic spike, resulting in activity-dependent decay (Eq. (17)).