François CholletVerified account

Humans are not insects; the few megabytes of brain-related information in your DNA define *how to learn* but contain relatively little *knowledge* -- and only animal-like basics. (reasons: can't pass on knowledge via DNA; not enough space; evolutionary scales not long enough)

Humans get to interact with the environment and get a lot more supervision. Try taking two random languages from a translation dataset that you don't speak and see how well you do against a neural network.

Another interesting case are medical imaging datasets if you don't have any background in medicine. Try learning to detect the classes in the NIH chest x-ray dataset from labels alone.

Do you think that a different "algorithm" or connection model will be needed for concept formation? It would seem that there is some mental tool that human brains have that animals do not, and just scaling won't bridge the gap.

I can show my 2yr 1mo old son some new object and name it both for a single time, and he will be instantly able (after seconds of processing) to show you similar objects and name them at 80+ accuracy. For example "steering wheel" (from ships). How could we mimic that in ML?