Figure 1. Schematic illustration of a recurrent neural network.

The network receives time-dependent input $\mathbf{𝐱}(t)$, and its synaptic weights are trained so that the output $\mathbf{𝐲}(t)$ matches a target function ${\mathbf{𝐲}}^{*}(t)$. The projection of the error $\mathit{𝜺}(t)$ with feedback weights is used for learning the input weights and recurrent weights.