AI-powered
podcast player
Listen to all your favourite podcasts with AI-powered features
Using a Linear Kaleidoscopy to Improve Neural Network Performance
Each layer of a neral network contributes a new set of hyperplanes, and the reus act to toggle the hyperplanes in an input sensitive way. Randall commented that in high dimensional spaces, you can easily separate everything, but your generalization performance might be bad. He thinks the merit of deep neural networks is being able to find a non linear transformation which retains separating hyperplanes while reducing dimensionality enough to confer generalization power.