your GAN is secretly an energy-based model

As I was reading this NeurIPS 2020 paper by Che et al., and trying to make sense of it, I came across a citation to our paper Casella, Robert and Wells (2004) on a generalized accept-reject sampling scheme where the proposal changes at each simulation that sounds surprising if appreciated! But after checking this paper also appears as the first reference on the Wikipedia page for rejection sampling, which makes me wonder if many actually read it. (On the side, we mostly wrote this paper on a drive from Baltimore to Ithaca, after JSM 1999.)

“We provide more evidence that it is beneficial to sample from the energy-based model defined both by the generator and the discriminator instead of from the generator only.”

The paper seems to propose a post-processing of the generator output by a GAN, generating from the mixture of both generator and discriminator, via a (unscented) Langevin algorithm. The core idea is that, if p(.) is the true data generating process, g(.) the estimated generator and d(.) the discriminator, then

p(x) ≈ p⁰(x)∝g(x) exp(d(x))

(The approximation would be exact the discriminator optimal.) The authors work with the latent z’s, in the GAN meaning that generating pseudo-data x from g means taking a deterministic transform of z, x=G(z). When considering the above p⁰, a generation from p⁰ can be seen as accept-reject with acceptance probability proportional to exp[d{G(z)}]. (On the side, Lemma 1 is the standard validation for accept-reject sampling schemes.)

Reading this paper made me realise how much the field had evolved since my previous GAN related read. With directions like Metropolis-Hastings GANs and Wasserstein GANs. (And I noticed a “broader impact” section past the conclusion section about possible misuses with societal consequences, which is a new requirement for NeurIPS publications.)

plenary talks at JSM 2017 in Baltimore

Poisson process model for Monte Carlo methods

“Taken together this view of Monte Carlo simulation as a maximization problem is a promising direction, because it connects Monte Carlo research with the literature on optimization.”

Chris Maddison arXived today a paper on the use of Poisson processes in Monte Carlo simulation. based on the so-called Gumbel-max trick, which amounts to add to the log-probabilities log p(i) of the discrete target, iid Gumbel variables, and to take the argmax as the result of the simulation. A neat trick as it does not require the probability distribution to be normalised. And as indicated in the above quote to relate simulation and optimisation. The generalisation considered here replaces the iid Gumbel variates by a Gumbel process, which is constructed as an “exponential race”, i.e., a Poisson process with an exponential auxiliary variable. The underlying variates can be generated from a substitute density, à la accept-reject, which means this alternative bounds the true target.  As illustrated in the plot above.

The paper discusses two implementations of the principle found in an earlier NIPS 2014 paper [paper that contains most of the novelty about this method], one that refines the partition and the associated choice of proposals, and another one that exploits a branch-and-bound tree structure to optimise the Gumbel process. With apparently higher performances. Overall, I wonder at the applicability of the approach because of the accept-reject structure: it seems unlikely to apply to high dimensional problems.

While this is quite exciting, I find it surprising that this paper completely omits references to Brian Ripley’s considerable input on simulation and point processes. As well as the relevant Geyer and Møller (1994). (I am obviously extremely pleased to see that our 2004 paper with George Casella and Marty Wells is quoted there. We had written this paper in Cornell, a few years earlier, right after the 1999 JSM in Baltimore, but it has hardly been mentioned since then!)