[silvertree]

Hello all-

I have a really simple question about Hebbian learning, but one that has been frustrating me for a while. If you could shed light on it I would appreciate it.

I really don't get how Hebbian learning works.

If there is an *incorrect* connection between node A and B, when A fires, B will fire, increasing the connection between A and B even though it was the wrong connection. Right?

Why is it a good idea to increase the link between A and B, when it is the link between A and B that is causing the correlation?

I could understand that this would be a good idea if A and B were independently being stimulated (e.g., if Bell, then Food)...but that is not the case with Hebbian learning, or am I getting something wrong?

I assume I am missing something critical here. Is it the fact that you look at whether A and B are both firing before propagating the signal from A to B (so, in effect, you are looking to see if A and B are indeed firing independently of the link from A to B)?

[flashprogram]

I believe it is the case that the strength of A alone is not usually sufficient to activate B. It is usually only when there are several inputs to B firing within a short period of time that B fires.

Thus A will usually only be able to activate B and thus become stronger, if it coincides with some other neurons acting on B(all the neurons involved will be strengthened too, if I recall correctly.).

[silvertree]

Thanks. That makes sense. But what about the case when the weight between A and B is strong enough such that every time A fires, B fires? Hebbian learning would continuously increase the weight of this link (even if that association is incorrect/spurious) and eventually max it out.

This seems like something that would happen a lot.

For example, if in environment X, A does frequently precede B, then Hebbian Learning would correctly produce a connection between the two.

But if the environment changed (say, to Y) and A and B are no longer correlated, how would Hebbian learning ever learn to forget the association?

Even with a decay rate for all links, as long as A is perceived frequently (even if B is never perceived), A will keep causing B to fire, and keep that connection strong.

Is that right? That seems like a pretty big problem to me.

[flashprogram]

I believe something must limit the strength of connections per neuron, under normal conditions, such that there is a need for the input of many neurons as a requisite for optimal firing of B.

I say this, because in the visual system you see that neurons seem to respond best to groups of incoming connections, not to individual connections, If I'm not mistaken(since the receptive fields become larger and response is optimal for more complex stimuli at later stages than at earlier stages. This site: http://hubel.med.harvard.edu/bcontex.htm, explains it better.).

Hebbian learning cannot be used to forget, as far as I know. There is something called long term depression that weakens connections. I will have to verify, but I believe that there is a competition between synapses, such that if A is consistently inactive at the same time other neurons cause B to fire, the connection of A is weakened while the others are strengthened.