Biologically plausible learning

[tfiers]

(I'm copying this here from the discord so it's more easily discoverable and linkable).

@KarimHabashy:

Just a quick question: I know STDP methods have a large number of variables to tune for learning. Is this the only reason for not having it mentioned beside gradient decent for learning and optimization?

@Helena-Yuhan-Liu:

I think that's a good question. A short response is that it's very hard for STDP to steer weight updates such that the network as a whole can achieve a certain goal. STDP may work fine if you have just a single layer. But having more layers necessitates hierarchical credit assignment, which backpropagation offers a solution to. That's why researchers have turned to backprop as an inspiration and propose biologically plausible approximations to hypothesize how the brain might learn. If you are interested, please check out Deep Learning Framework for Neuroscience by Richards et al 2019. They explain this very well.

[Helena-Yuhan-Liu]

Karim's follow up question: "1) There is progress in training with hierarchical networks, like the burst dependent plasticity by Payeur, is there barriers for considering this? 2) Depending on the input format, maybe, the localization can be learned in a 1 layer with careful choice of the input parameters (like axonal delays)? I mean an argument can be that sound localization is a mental function and there must be multiple layers between the cochlea and the higher brain areas?"

For 1), I agree there has been great progress in biologically plausible learning rules for hierarchical networks (and recently RNNs as well), such as the BurstProp example you mentioned. One barrier is that these gradient approximations may require more training iterations. The intuition is that if you look at Figure 3 in Richards et al 2019, you will see that partially following the gradient necessitates a longer path to the objective. Additionally, these rules can sometimes be cumbersome to implement compared to gradient descent learning using automatic differentiation.

On a slightly tangential note to your question, whether you should use bio-plausible rules vs backprop also depends on if your question is concerning how to reach a (local) optimum, or if you just care about what a (local) optimum can be and don't care about the how. If your question is about the how, e.g. how does a network learn to solve a delayed match to sample task, then bio-plausibility of your rule is essential. If your question is about the what, e.g. what could be a connectivity structure that give rise to the pattern you are interested, then doing backprop to find a (local) optimum could be a lot easier.

As for 2), I have not really worked with sound localization so I don't feel qualified to answer. I am definitely interested in hearing the response from someone.

[thesamovar]

Another reason to go for gradient stuff is that STDP has been well studied, including for sound localisation. It would be more interesting to try another approach that hasn't yet been tried!