Deep Learning, learning completion, silicon chip implementation

[kris kaczmarczyk]

TL;DR Summary

Back propagation, learning process, coefficients at the hidden layer nodes. Hard-coded on the silicon?

After lengthy process of "deep learning" and back propagation of information; would we get static coefficients for the thousand of nodes; for example for playing game of chess, would that state be good to hard-wire on the silicon cheep and this way we would have perfect chess player?

Or the coefficients are always changing? In other words can we stop learning at certain stage of error percentage and than we have some set of numbers which we can hard-code on to the hardware (driving cars, translating, painting)?

https://www.popularmechanics.com/technology/design/a28816626/worlds-largest-computer-chip/

 

[BvU]

https://en.wikipedia.org/wiki/Solving_chess#Predictions_on_when/if_chess_will_be_solved

two trillion (2e12 since the article is inch-based it must come from one of those countries) transistors sill isn't much if you have to deal with 1e43 board positions...

But it'll be a nice step forward

You would need a lot of those chips ! I wonder if anyone can make a Shannon-like guess ?

Note that one layer all over the surface of the world only gets you 1e16 chips