Gary Marcus

Great! New thread then: What are the ways in which people in the neural net community think that today's ANN's are NOT perfect models of the brain? Where do you want to dig deeper, and why?https://twitter.com/timkietzmann/status/1200159447651438592 …

Not a neuroscientist, but my own opinion is that weighted gating, compositionality, and learning are the foundations of bio neural nets, and all other phenomena that we care about are built on these blocks. Traditional computers only do the first, but DL can do all three.

I personally see no reason a priori why the specific mechanism of these three functions (kludged together by incremental evolutionary silliness) would matter to the emergent phenomena, so arguments like "neurons don't back-propagation" seem a bit spurious to me.

I realise I'm outing myself as a connectionist here, but I personally feel that trying to model the more interesting feedback loops in the brain with existing building blocks will be more beneficial than trying out different paradigms.

What I think other approaches (eg explicit symbol manipulation) offer is shortcuts. Humans know things, so it would be more efficient to embed things we know into a system (vs learning everything). But I don't think we need it, and it has always been more brittle than learning.