Input Feature Combination

I'm not sure when I figure it out, but the two layer mechanism is just getting co-occurence measurements.

Combination by Timing

It's well known you can't get xor with this, so I tried to modify the system to get it.
It turns out you can get it by changing the timing of presenation.
In retrospect, I don't think this is really a good way forward.
The timing is a bit arbitrary (not A 10ms, A 11ms ... category 15ms).

Still there are lots of data sets where we need feature combination.
A standard one is the monks from the UCI benchmark.
We added an extra layer to combine feature pairs.
In the task we were interested in there were 6 features, so we made neurons for the inputs, and for each pair.
So, there was a neuron for feature 1 value 1 and feature 2 value 1.
The synaptic weights were static and set so these feature combination neurons fired only when both input feature value pairs were present.
We set up a statistical system that categorised entirely by co-variance of singles and pairs, and the STDP system got the same result.

There's no reason this mechanism can't be used to combine multiple features.

It might be interesting to use learning instead of static synapses.

Timing could be an issue as the combination layer necessarily fires after the input, and has to fire before the output to learn properly.