
Embedded Machine Learning: Part 2 (Ep. 183)
Data Science at Home
00:00
How to Prune a Network and Find a Winning Ticket
The pruning is not a destructive pruning, but you just create a mask that will, you know, mask these parometers out of the computation from the nextiteration onward. And in the fourth and last step, very simply, you reset the remaining parometers to their initial value inteta zero - which was the random, initialized parometers at iteration zero. First iteration, and that's your winning ticket. So this is a very systematic way of finding while, first of all, of pruning a network and finding, or preserving, finding a winning ticket whilst preserving accuracy.
Transcript
Play full episode