SNN's and learning (some thoughts on)

(1) What is input, what is output ?

When you use a classical Backpropagation Network, you have one layer of neurons for input. and another one for output. Same it is in nature, when it comes to sense organs for the input and nerves to the muscles as the output. Same in robotics.

But what, when the input is of ideas and the expected output is of the same type ?
Everything happens in the mind then, without direct relation to the real physical world.

In such a case, input and output are patterns, to be more specific: A (current) state of the entire brain.

(2) Initial optimization of topology

The neuron model should be transformed into an euklidean space by using an algorithm, that can optimize the graph for shortest links (axons).

Nature did this over million years during evolution. When we start with an empty brain and no evolution at all, we should simply 'sort' it for shortest links.

This step also groups the neurons by meaning in space, e.g. groups synonym words so that they become neighbours.

(3) Working with Brain-Patterns

When thinking about making something useful out of SNN's, then categorization (grouping, clustering) of data comes first into my mind. E.g. a SPAM filter software that can categorize emails and so detect spam.

This applicaton has these components:

* the neuron network, that has run for a while and has built up new connections by playing the ancient game of 'what fires together wires together'.
* Input data that is used to trigger some neurons. This leads to a 'brain state' after some iterations.
* Comparison data that was first run same as 'input data', in order to compare both patterns to see, how much they have in common, how much the rate of convergence is.

(4) Representing brain patterns

Biological brains don't store their pattern states at all. The current state is the current pattern, that's all.

When using computers, you should look for imaging algorithms with data reduction / compression.
You could represent your brain's state by using wavelets and DWT. This also has the advantage, that you can process it with different level of details.

You then build an algorithm that calculates rate of convergence for two patterns.

This is all about mind. When you build roboters or you want to control machines, there should be some neurons that do the job and fire their signal to an output. Senors are connected with neurons that fire, when the sensor signal is above threshold.