Introduction

This episode's guest is jonathan frankl who as a pe strende. He talks about the lotradicat hypothesis, and some of the foll work that's come out after the posision pabe was written. The experiment i wanted to try was maybe these connections received a very lucky initialization. Maybe the original values that each connection received were very important. We hypothesise perhaps, in general, the large networks that we train have much smaller sub networks that are capable of training just as effectively. They can train just as to just as good a performance. But t was a way of summarizing all our observations on a pretty range of vision networks.