Kohonen Networks
After researching Kohonen networks online, I've figured out how to train them and such. But if anyone here could help me with my biggest questions, I'd highly appreciate it.
1) What are the outputs of the network given an input vector? Is it the position of the winning neuron in the neuron matrix? If not, how do I calculate the output?
2) Are the neurons in the neuron matrix weighted to each other? If so, how they affect each other? (The only effects I've seen neuron matrix neurons have on each other are their proximity in the neuron matrix and grouping of similar weight vectors to the input neurons)
Thanks in advance!
[654 byte] By [
Squealsa] at [2007-10-1 3:13:29]
> After researching Kohonen networks online, I've
> figured out how to train them and such. But if
> anyone here could help me with my biggest questions,
> I'd highly appreciate it.
>
> 1) What are the outputs of the network given an input
> vector? Is it the position of the winning neuron in
> the neuron matrix?
Yep, that's right.
> 2) Are the neurons in the neuron matrix weighted to
> each other? If so, how they affect each other? (The
> only effects I've seen neuron matrix neurons have on
> each other are their proximity in the neuron matrix
> and grouping of similar weight vectors to the input
> neurons)
I'm not sure what you mean by weighed to each other, but there is a relationship between neighboring neurons in the network. That is because when you train them, you turn the "winning" neuron towards the training vector, and the neighbors also get turned slightly. So in the end, neighboring neurons will point in a similar direction.
> Thanks in advance!
On another note, Kohnonen networks are conceptually "cool", but the problem they solve is the least squares classification problem, which is solved by other, faster algorithms like k-means. I'm not sure if your interests lie in neural networks or if your looking for a classifier, but if its the latter, there may be better thechniques available.
Good luck.
