In plain words: A new kind of generative network learns step by step, adding new features—like physical symmetries—on top of old ones, and pairs with a classifier to also sort images. On handwritten digits it beat similar networks at both sorting and estimating image likelihood.
Abstract
We study from a physics viewpoint a class of generative neural nets, Gibbs machines, designed for gradual learning. While including variational auto-encoders, they offer a broader universal platform for incrementally adding newly learned features, including physical symmetries. Their direct connection to statistical physics and information geometry is established. A variational Pythagorean theorem justifies invoking the exponential/Gibbs class of probabilities for creating brand new objects. Combining these nets with classifiers, gives rise to a brand of universal generative neural nets - stochastic auto-classifier-encoders (ACE). ACE have state-of-the-art performance in their class, both for classification and density estimation for the MNIST data set.
Galin Georgiev
arXiv:1508.06585 · cs.CV, cs.LG, cs.NE · submitted Aug 26, 2015 · updated Jun 30, 2016
abstract · pdf · html · v5: added thermodynamic identities and variational error estimation; expanded references
- Doesn't get to the point until about halfway through
- Repeatedly mentions Einstein
- Appeals to quantum mechanics out of nowhere
The style is particularly hard to parse (or I'm particularly dense, but I am generally comfortable reading papers on variational inference, neural networks, etc). At the same time, a lot of it rings true-ish... Does anyone get where they're going with this?