Gary Marcus

of course, a lot rests on what you mean by right architecture, and whether you include symbol-manipulation (which we argue is a key component) in that scope. you and i probably ultimately agree on the value of *hybrid* architecture.

I think we can agree that the brain has specialized components that work in concert to achieve human cognition. We can furthermore agree that each of these components is built of specialized neural networks. Hybrid architectures are a stepping stone towards more general kinds.

if you have specialized components that operate under different principles, you have hybrid systems.

Neuroscientists will unanimously agree that every specialized component of the brain is made of up of neural networks.

of course. as i have written numerous times, the question is not whether the brain is a neural network, but rather what *kinds* of neural networks the brain uses, whether they have anything to do with the tools in Ml, and what sorts of computations they implement.

Indeed. E.g. we cannot solve intelligence under the feedforward networks paradigm which seems particularly suited for motor control but not for reasoning and general problem solving. NNs with supervised learning based on 'output' values are nowhere to be seen in the brain

oh, i think the brain does some supervised learning, even if we aren’t sure yet of the mechanisms.

Agree, supervision is present, but how? Kids don't create a 'dog's representation by hearing the word in the presence of the animal: they associate the label to a representation they *already had created* by other mechanisms; plus, supervised information comes also in the *input*