Comment

Comment by Álvaro Calle Cordón on May 31, 2016 at 10:05pm

Yes, I agree with Lars, you won't probably need to dive too deep into physics analogies and spin systems to have a first glimpse on Boltzmann machines. I you have no formal education in Physics this may be overwhelming somehow. I just wanted to point out where the main idea came from. 

Comment by Lars Schwabe on May 31, 2016 at 1:51pm

If you want to know about Boltzman machines from a very fundamental point of view, I suggest to read the original papers, i. e. go back in time a bit. Hinton, one inventor, is *not* a hard-core math person, so the early papers contain quite some informative text and justification and intuition. Personally, and I did PhD with a physics professor within computer science, I dont find the physical analogies so helpful, though. I dont need to dive deep into spin systems. The formal properties of such systems are sufficient for me.

Comment by Álvaro Calle Cordón on May 26, 2016 at 10:51pm

I think these are advanced machine learning methods. If you want to understand the basis of Boltzmann machines you might want to learn about the Ising model as a first step. The Ising model was originally introduced in Statistical Mechanics (Physics) to model the interaction of a ferromagnetic system where the interaction between particles is given by the spin-spin interaction. The Energy in the Boltzman machine is nothing else but a spin-spin interaction like in the Ising model. You can see the spin as a property of a particle (or a unit in the network) which can take two states (in physics +1 or -1 and in AI 0 or 1). The probability is given by the Boltzmann distribution with this interaction energy term.