Jason Antic

We don't have to know everything but he's right that we don't even know enough to get to say "this is an abstraction"; you can't abstract if you don't know what's important. I've worked with spiking nets, biological nets, and DL

And if for the spiking nets I could carefully say that some properties of in vitro networks are reproduced, I wouldn't say it's an abstraction of the brain for example. For DL it's not even extremely controversial to say that they're not an abstraction of biological networks.

It's a goal in neuroscience to build compact mathematical models of individual neurons that accurately reflect how neurons in the brain work. I think Gary's point is we don't know how neurons in the brain work, so it's hard to assess how good the models we have today are.

you don't have to know everything, but given we can't answer basic questions, eg -what is computational role of dendrites? -why do synapses have 100s of distinct proteins? -why roughly 1000 cell types? -how is short-term memory neurally realized? we can't be confident of much.

Thanks Gary! I do appreciate you trying to cut through the hype in the field of machine learning- it’s definitely needed.

I was a neuroscientist, and I can assure you it's not unfair criticism. We can say that they're inspired by part of the workings of a class of neural networks, but take care of people who try the analogy too far. Specially people who think they can correct psychology.

Let's make a list of mechanisms not currently addressed by NNs (note that they don't need to, just to emphasize the differences): * Spiking neurons * Second messengers (maybe LSTM, but even then I doubt the analogy is complete) * Oscillating recurrent circuits (with chaotic dyns)