Learning and retrieval

In classical ``Hebbian'' studies on single-population recurrent networks, connection weights are set according to a given set of pre-defined sequences [19,15], without dynamical interaction. On the contrary, our system uses an on-line learning rule, so that the weights are updated at each time step during the learning process. Weight adaptation thus grounds on a real-time interaction between the input signal and the system dynamics. In order to evaluate the effect of the learning rule on the dynamics, we alternate learning phases (Eq.(2)) and testing phases (Eq.(1)) A first attempt to learn sequential patterns of activation from a background chaotic activity with an on-line Hebbian learning can be found in [18]. Otherwise, a Hebbian learning rule has been proposed on our model in [10] for the dynamical encoding of static input patterns. For simulations on the ReST model, learning takes place on inner and feedback links, i.e. $\varepsilon ^{(12)}=0.1$ and $\varepsilon ^{(22)}=0.02$, so that the learning ``strength'' is lighter on secondary layer recurrent links. Quantitative effects of parameter $\varepsilon ^{(22)}$ on the learning capacity can be found in section 4.3. Otherwise, we have $\varepsilon^{(11)}=\varepsilon^{(21)}=0$ (the weights from the primary layer remain unchanged).