Perturbed gradient descent with backprop is what we were doing in the 90s. It feels like there are some new tricks but mostly what I see is the result of GPUs and cheap memory.

I volunteer at a summer science camp and 90% of the projects are "I pointed AI at this problem and...", nobody seems to be even trying for analytical approaches any more. I'm ready for a new fad.

Anybody else remember when it was all wavelets all the time? That was kinda fun.