Figure 6.

The standard (CD)_k procedure, compared to event-driven CD. (A) In standard CD, learning proceeds iteratively by sampling in “construction” and “reconstruction” phases (Hinton, 2002), which is impractical in a continuous-time dynamical system. (B) We propose a spiking neural sampling architecture that folds these updates on a continuous time dimension through the recurrent activity of the network. The synaptic weight update follows a STDP rule modulated by a zero mean signal g(t). This signal switches the behavior of the synapse from LTP to LTD, and partitions the training into two phases analogous to those of the original CD rule. The spikes cause microscopic weight modifications, which on average behave as the macroscopic CD weight update. For this reason, the learning rule is referred to as event-driven CD.