Post #3195790

2025-01-23 17:12 UTC

@GabrielBena@neuromatch.social Our intuition suggested that resource constraints are likely to be important: there's little incentive to specialise if you have infinite resources. Sure enough, when we did large parameter sweeps we see that you get more specialisation when resources (neurons, synapses) are tight. This seems like an important insight: we see that resource constrained biological brains are great at generalisation, an expected outcome of having specialised modules with generalisable functions, while machine learning systems are not. Maybe we give them too much computational power? 🤯

Replies (2)

@neuralreckoning@neuromatch.social 2025-01-23 17:12

@GabrielBena@neuromatch.social Finally, we made use of the fact that our networks are recurrent (which was a necessary restriction of the simple architecture that we used) and checked how specialisation changed over time. Intriguingly, it decreases over time. But there's more. This drop in specialisation happens faster the more synapses between the modules and the less noise there is. This looks like maybe specialisation falls simply as a result of how much net communication bandwidth there is between the modules. This raises the question: maybe specialisation isn't as simple as we think. Perhaps to some extent it's just a measurement artifact of limited communication bandwidth? Or maybe, understanding information flow is key to building systems that can specialise and generalise? These are the sorts of questions we're following up on now, hopefully we'll have more to say about that soon, but in the meantime we'd love to discuss some of the issues and questions raised with you all. What do you think?

Open ##3195791
@m8ta@fediscience.org 2025-01-23 20:17

@neuralreckoning@neuromatch.social @GabrielBena@neuromatch.social Thanks for sharing, solid work. Some thoughts on this point: "we see that resource constrained biological brains are great at generalisation, an expected outcome of having specialised modules with generalisable functions, while machine learning systems are not. Maybe we give them too much computational power?" There seem to be \approx zero generalization penalties from overparameterization in networks trained via SGD. Eg: train a 4-layer MLP on MNIST. Performance is identical if the two hidden layers are anywhere from 128 - 16k wide. Is it true that resource constraints engender better generalization in biological networks? (Trained, presumably, without SGD?). I see no plots or metrics of generalization - ? Regarding functional specialization, imho how networks learn is a interplay between 'natural' spectral / eigenmode / Fourier-coefficient learning (axis-agnostic), and more interpretable axis-aligned learning, as driven by asymmetries in e.g. Adam or regularizers (or biology). Again, in practice forcing functional specialization tends to *negatively* impact ANN generalization performance; instead, it seems to be primarily a function the networks structure, and not of the quantity of gross computation therein (above a limit). The former which, of course, you've varied, in an interesting way... so where's the generalization figure? :-)

Open ##3195796