Figure 4.

The different studied architectures and three-factor STDP. (A) Illustration of our main model (Equations 31, 32) with its two sets of synapses (black) and (gray). The “error signal” e_N(t) modulates the learning of the -synapses. (B) Illustration of the simplified model (Equations 33, 34) with its single set of synapses . The “error signal” e_N(t) modulates the learning of the -synapses. (C) Learning in the main model's -synapses is triggered by the covariance between the Hebbian trace H_ij and the moving average of free energy . The inset blue curve indicates the shape of the Hebbian trace term as a function of the time interval between pre and post synaptic spikes. The dotted gray line corresponds to H_ij = 0 (no weight change). A positive covariance will induce synaptic depression (left quadrant of blue curve) while a negative covariance will induce synaptic potentiation (right quadrant of blue curve). A null covariance would induce a centered random walk on the synaptic weights.