Diehl and Cook, and Rybka

I was trying to get some spiking net STDP learning working because a lot of project students are interested in doing theses in ML.
A few years back I ran across Unsupervised learning of digit recognition using spike-timing-dependent plasticty (Diehl and Cook 2015), and said, hey I should try this out.
I spent a fair bit of time (80 hours?) at it, but didn't really make much progress. It turns out placing a bunch of neurons in four dimensional space does better than chance.
I was then at a (virtual) conference and Rybka presented some results. I said, send me the code. He did. I unpacked it. Then modified it for the two tasks below.
There are inputs. In the Iris case there are four features, so there are four neurons, which are really just spike sources.
The categorisation neurons are like the SOM nodes.
When they fire they each turn on their own inhibition neuron, which inhibits the rest of the categorisation neurons.
That's the winner take all network.
The synapses from input to categorisation are plastic.
So, present an input for a while, have a rest period, then present the next and so forth.
One of the cool things is the neuron with dynamic threshold and the full mechanism automatically normalizes.