Extensions

This may be a bit specific, but we used multiplicative weight changes in the STDP rule instead of additive.
This kept the synaptic weights from saturating.
We also used a training timing mechanism where incorrect categories were presented before the input, the input, then the correct category.
With STDP this meant that the weights went down for the "bad" category and up for the "good" one.
Note that this is a really an artificial use of time.
Just to dot the I-s, we present in epochs with one training instance followed later by another. The double category presentation mechanism seems to make this time less important for reducing weights.