Deep Learning in a nutshell

Akstually.. no.. unless you dont use any activation function other than linear.

@purist yeah, it was an oversimplification for a meme, ofc the multiple layers alone are enough to classify it as not “just a matrix”, but the fact that the first attempt/core idea was born from a matrix blew my mind the first time I approached it! A bit like when I first heard about matrix on school back when I was 16 and immediately thought about mmorpg’s inventories! (I know they are actually arrays, but as soon as I saw it with no coding experience I was like: “they must be using something like this for video games, except they store the id, the amount and maybe the picture” ... close call!)

At least, optimizing weight matrices is sort of, what training a neural network actually does...
So given a model, the "intelligence" is in a set of matrices of weights.

Therefore "Wait, deep learning is only about optimizing matrices?" would not be wrong.

Some are matrices, some are graphs, most are really long chains of conditionals sitting behind an overzealous sales team