Natural Kinds

All of those things worked with nets where the neurons were activated by the environment.
If we're really going to get neurons to learn the way we do, the neurons have to fire on their own.
We changed our model so that hypo-fatigued neurons fired.
We could then move into neurons that weren't externally stimulated.
We could then use topologies like:
We used one like this and one with four subnets to categorise Irises, Cars and Yeast, again from the UCI database.
We presented input neurons and the correct category during training.
We measured either by counting the spikes in the answer neurons or comparing training nets firing behaviours to test nets using Pearson measurements.
Pearson did better, but both had reasonable performance compared to standard algorithms.
Note, all of these tasks are natural kinds. They don't have boolean operations.
The xor problems still pop up. We can do it, but we have to throw a lot of neurons at the problem, and kind of know it's coming.