Reviews: Learning Infinite RBMs with Frank-Wolfe

This paper is well-written and addresses the important practical issue of choosing the number of hidden nodes in an RBM. The application of the Frank-Wolfe algorithm is not particularly novel, given the large set of related work that the paper cites. But the paper makes the key technical move of replacing the discrete sum over hidden nodes in the marginal likelihood function with an arbitrary distribution over weight vectors. This opens the path to applying Frank-Wolfe in a functional gradient setting. I found this move interesting and novel (although there may be related work I don't know about); it might suggest similar approaches for other undirected models besides RBMs.