Gary Marcus

There's a major difference: DL was guided in its infancy by ideas from neuroscience, so there is a relatively direct link between them. In contrast, the application of quantum mechanics to the c-word is taking two distinct fields and tying them together on speculative grounds.

The relation between deep learning - with its single neuron type and largely homogenous architecture - and the actual complexity of the human brain, with > 1000 neuron types, hundreds of proteins at each synapse and > 100 distinct brain regions - is risible.

Every model is an abstraction. Newtonian mechanics ignores air turbulence, molecular interactions, etc. Climate models capture coarse grained interactions, not the multitude of animals, plants, and wind-currents that truly shape the climate. Neural networks are no different.

Let's be real. Current neural nets have been shown empirically to work on some problems (after tinkering to get details right) - but do we really *know* that they are an abstraction of the brain, in which their details map onto simplifications of actual brain processes? No.

I'm sorry, but this is a bad take. Yes, we know they are simplifications of real brains. 1) Neurons do something very similar to linear integration with a non-linearity. 2) They process inputs in a distributed, parallel manner. ANNs capture this basic process, period.

ANNs are almost entirely feedforward, though. We don't really have a good ANN abstraction of how feedback processing plays its part in the brain.

No. ANNs are not entirely feed forward, most advanced models now incorporate recurrence.

And those recurrent models are nowhere near as well-studied and easy to engineer as the feed forward ones, nor do we know that their precise patterns of recurrence resemble the brain, rather than being useful engineering heuristics.