Episode 32: Jamie Simon, UC Berkeley: On theoretical principles for how neural networks learn and generalize
Generally Intelligent
00:00
Percolation Theory and the Lost Landscape
There's a path through this hydrometional space. And actually, these paths are really simple. You can usually find a path that only has one kink. It's just piecewise linear with two lines. But it seemed to me like clearly this is a little of high dimensionality. Just like with the fact that they're no bad local minima and high few dimensions well if you have a million them choose from probably you'll find a path from point A to point B err. We'll find a piecewise path where only a small fraction of samples are ever misclassified along the wall. This was interesting to me because they appealed a very u. I was like obsessed with
Transcript
Play full episode
