A lot of my close friends have left physics and become enamoured with machine learning. I have always felt there is an unsatisfactory level of emphasis placed on understanding what is going on. It was amusing to note that after a great colloquium at McMaster today by Roger Melko on applying machine learning to the many body problem that the immediate response of the senior faculty was exactly the same.
I was inspired by this paper which uses principal component analysis to learn about a generalized Ising model. I asked Roger about it and he pointed out that PCA does not scale as well as neural networks, but the latter returns an incomprehensible, and often unphysical, description of the problem at hand.
It would be really interesting if someone could develop a kind of generalized PCA to apply to a scalable solution so as to “rotate” the weights and biases of a neural network into a kind of “basis” and identify the principle axes.