Gary Marcus

actually, convolution (translation invariance) does not need to be built in.

i don't know a system that learns translational invariance, and i'd think you would want to have it. you could use a different starting point, though.

Isn't innate machinery an architecture search done through natural selection? Unless you believe in intelligent design, that machinery did emerge on its own through world interactions. Why not try to simulate this search process rather than hand designing the machinery?

in principle you could if you were patient enough; in practice, people start from near zero, and don't get that far In most domains of engineering we choose NOT to start with random search, w good reason Why should ML be different? Why ignore hints from products of evolution?

@GaryMarcus , assuming some sort of Pareto principle, if you had to pick three items of this list, which would they be?

operations over variables, causality, and capacity for some basic forms of representation. but my general point is that even those three probably wouldn’t be enough. our genomes have about 20k genes, most expressed in brain. why limit this list to 3?

Which ones are from Spelke?

if you want to be a pro, maybe read a bit of her work on "core knowledge"? it's crazy relevant.

I worry that this debate won’t be fruitful because we don’t have a commonly accepted definition of “intelligence”. Participants will reason based on their own assumptions of what it is, and will end up talking past each other, as usual.