[Article] Is Geoffrey Hinton predicting the emergence of “Supersymmetric Artificial Neural Networks”?
Source video: https://www.youtube.com/watch?v=rtGXv88UQ-c
The motivation behind Hinton’s recent capsule neural network can be summarized in one line from Hinton’s talk above at 23:11: “A powerful way to recognize things is to get two measured multi-dimensional structures to make predictions that ‘agree’ in high dimensional space”.
On a separate note, I’ve been trying to invent something called the “Supersymmetric artificial neural network”, that is in the SU(M|N) special unitary group family of geometries, based on evidence of supersymmetry in biological brains.
Supersymmetry can naturally yield “partner potential” signals beyond the phase/magnitude space scope of unitary or complex valued neural nets aka SO(n) — orthogonal or U(n) — unitary valued nets. (See this earlier answer of mine for a nice clear overview of geometric solutions observed in artificial neural network history)
I wonder if the SU(M|N) special unitary group based partner potential signals can yield the “agreement” scenario Hinton mentions at 23:11 of the video?
▽
▽
▽
Simple Supersymmetry explanation: https://www.youtube.com/watch?v=0CeLRrBAI60
So a supersymmetric configuration can be seen as a symmetric order over units of complex or sophisticated dimensions.
Could what Hinton refer to at 23:11 relate to some “swap and maintain order condition” over complicated regions involving supersymmetry?
▽
▽
▽
Quora version with math symbols is available here.
▼
▼
Author:
I am a casual body builder, and software engineer.
