Gary Marcus

Following on from a discussion with @utafrith and @CeliaHeyes, it's interesting to ask how much 'start-up kit' a brain (or an A.I.) needs to set it up to learn the right things in the right ways...

The answer depends on what we want it to do. If we want it capable of understanding natural language and performing abstract reasoning, we'll need more than deep learning...

But how much (and what bits) of brain architecture is devoted to those higher functions (versus supporting what we might call 'intelligent behavior' of animals having to survive in the real world)?

We can at least say that the start-up kit of a human brain supports those higher functions in a way that the start-up kit of a chimpanzee doesn't...

Of course, current A.I. doesn't get near even chimp level general intelligence (though it surpasses even human performance on some kinds of tasks)...

Current AI doesn't even get near *mouse* intelligence--and a great deal of what mice (and most other animals) do is innate. If we could get achieve artificial mouse intelligence, we'd have a good foundation for human intelligence.

I agree. At least if that is the type of artificial intelligence we are after - the kind that allows a wee beastie like a mouse to do all the cognitive operations necessary for mouse survival

And I would argue most of the brain is devoted to this kind of adaptive control of behavior. And it would be super cool to build AI like that - as artificial agents...