36 points by 082349872349872 4 days ago | 4 comments
Maro 13 hours ago
I'm a physicist, and it always bugged me that we use Hamiltonians/energy/interactions to define spin glasses. I wondered if it'd be possible to just use raw probabilities on a grid and get the same/similar behaviour (eg. phase transitions, long-range correlations, etc).
So I played around with this idea for quite a while by running MC simulations, and ended up writing these articles (on my blog) about it.
But beware (i) none of this is peer-reviewed and (ii) I'm not a practicing academic physicist [so I may have missed things that would be obvious to a practicing academic physicist].
If anybody is interested in this, hit me up, happy to speak and/or play with the topic again.
leumassuehtam 17 minutes ago
foehrenwald 14 hours ago
deepnet 16 hours ago
So this paper is worth reading if you study learning in neural nets
[0] Spin-glass models of neural networks Amit et al 1985 https://journals.aps.org/pra/abstract/10.1103/PhysRevA.32.10...
This was why Hopfield was awarded a physics Nobel - spin glasses as emergent learning.
Along with Expected Energy of Brownian motion in gases for Hinton’s Boltzmann Machines ( and of course backprop being Newtonian downhill on the manifold of correctness ).
Langland’s worthy crossovers and the utility of the generalisation.