If you think about it deep learning is the triumphant return of analog computing.
Conversation
“Digital” is etymologically related to counting on fingers
“Analogue” is etymologically related to seeing in patterns/proportions
The symbolic vs subsymbolic divide is superficial, the discrete vs continuous divide gets closer, but the essence is how you measure the real
2
2
21
Digital operations measure things in terms of countable relations to an abstract, absolute standard. Analog operations measure things in terms of similarity relative to each other.
3
2
9
Replying to
Not sure i agree - the core property of digital is free exact copy operation. Analogue is defined by costly lossy copy operation. Tensor computing, for all it's mushiness, is in the former category.
1
1
Replying to
At the application level of abstraction though, it is lossy copies, like lossy copy of Mozart music, or lossy deep fake
1
Replying to
As in, training on the same data set repeatedly would produce a slightly different function every time? Or what the lossy part is?
1
Replying to
It's... not supposed to? I'm still not following what makes it analogue. More precisely, i don't see a groove-is-sound kind of lossyness and/or degradation in there.

