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 realised 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.)

frontier of simulation-based inference

“This paper results from the Arthur M. Sackler Colloquium of the National Academy of Sciences, `The Science of Deep Learning,’ held March 13–14, 2019, at the National Academy of Sciences in Washington, DC.”

A paper by Kyle Cranmer, Johann Brehmer, and Gilles Louppe just appeared in PNAS on the frontier of simulation-based inference. Sounding more like a tribune than a research paper producing new input. Or at least like a review. Providing a quick introduction to simulators, inference, ABC. Stating the shortcomings of simulation-based inference as three-folded:

1. costly, since required a large number of simulated samples
2. loosing information through the use of insufficient summary statistics or poor non-parametric approximations of the sampling density.
3. wasteful as requiring new computational efforts for new datasets, primarily for ABC as learning the likelihood function (as a function of both the parameter θ and the data x) is only done once.

And the difficulties increase with the dimension of the data. While the points made above are correct, I want to note that ideally ABC (and Bayesian inference as a whole) only depends on a single dimension observation, which is the likelihood value. Or more practically that it only depends on the distance from the observed data to the simulated data. (Possibly the Wasserstein distance between the cdfs.) And that, somewhat unrealistically, that ABC could store the reference table once for all. Point 3 can also be debated in that the effort of learning an approximation can only be amortized when exactly the same model is re-employed with new data, which is likely in industrial applications but less in scientific investigations, I would think. About point 2, the paper misses part of the ABC literature on selecting summary statistics, e.g., the culling afforded by random forests ABC, or the earlier use of the score function in Martin et al. (2019).

The paper then makes a case for using machine-, active-, and deep-learning advances to overcome those blocks. Recouping other recent publications and talks (like Dennis on One World ABC’minar!). Once again presenting machine-learning techniques such as normalizing flows as more efficient than traditional non-parametric estimators. Of which I remain unconvinced without deeper arguments [than the repeated mention of powerful machine-learning techniques] on the convergence rates of these estimators (rather than extolling the super-powers of neural nets).

“A classifier is trained using supervised learning to discriminate two sets of data, although in this case both sets come from the simulator and are generated for different parameter points θ⁰ and θ¹. The classifier output function can be converted into an approximation of the likelihood ratio between θ⁰ and θ¹ (…) learning the likelihood or posterior is an unsupervised learning problem, whereas estimating the likelihood ratio through a classifier is an example of supervised learning and often a simpler task.”

The above comment is highly connected to the approach set by Geyer in 1994 and expanded in Gutmann and Hyvärinen in 2012. Interestingly, at least from my narrow statistician viewpoint!, the discussion about using these different types of approximation to the likelihood and hence to the resulting Bayesian inference never engages into a quantification of the approximation or even broaches upon the potential for inconsistent inference unlocked by using fake likelihoods. While insisting on the information loss brought by using summary statistics.

“Can the outcome be trusted in the presence of imperfections such as limited sample size, insufficient network capacity, or inefficient optimization?”

Interestingly [the more because the paper is classified as statistics] the above shows that the statistical question is set instead in terms of numerical error(s). With proposals to address it ranging from (unrealistic) parametric bootstrap to some forms of GANs.

mining gold [ABC in PNAS]

Johann Brehmer and co-authors have just published a paper in PNAS entitled “Mining gold from implicit models to improve likelihood-free inference”. (Besides the pun about mining gold, the paper also involves techniques named RASCAL and SCANDAL, respectively! For Ratio And SCore Approximate Likelihood ratio and SCore-Augmented Neural Density Approximates Likelihood.) This setup is not ABC per se in that their simulator is used both to generate training data and construct a tractable surrogate model. Exploiting Geyer’s (1994) classification trick of expressing the likelihood ratio as the optimal classification ratio when facing two equal-size samples from one density and the other.

“For all these inference strategies, the augmented data is particularly powerful for enhancing the power of simulation-based inference for small changes in the parameter θ.”

Brehmer et al. argue that “the most important novel contribution that differentiates our work from the existing methods is the observation that additional information can be extracted from the simulator, and the development of loss functions that allow us to use this “augmented” data to more efficiently learn surrogates for the likelihood function.” Rather than starting from a statistical model, they also seem to use a scientific simulator made of multiple layers of latent variables z, where

x=F⁰(u⁰,z¹,θ), z¹=G¹(u¹,z²), z²=G¹(u²,z³), …

although they also call the marginal of x, p(x|θ), an (intractable) likelihood.

“The integral of the log is not the log of the integral!”

The central notion behind the improvement is a form of Rao-Blackwellisation, exploiting the simulated z‘s. Joint score functions and joint likelihood ratios are then available. Ignoring biases, the authors demonstrate that the closest approximation to the joint likelihood ratio and the joint score function that only depends on x is the actual likelihood ratio and the actual score function, respectively. Which sounds like an older EM result, except that the roles of estimate and target quantity are somehow inverted: one is approximating the marginal with the joint, while the marginal is the “best” approximation of the joint. But in the implementation of the method, an estimate of the (observed and intractable) likelihood ratio is indeed produced towards minimising an empirical loss based on two simulated samples. Learning this estimate ê(x) then allows one to use it for the actual data. It however requires fitting a new ê(x) for each pair of parameters. Providing as well an estimator of the likelihood p(x|θ). (Hence the SCANDAL!!!) A second type of approximation of the likelihood starts from the approximate value of the likelihood p(x|θ⁰) at a fixed value θ⁰ and expands it locally as an exponential family shift, with the score t(x|θ⁰) as sufficient statistic.

I find the paper definitely interesting even though it requires the representation of the (true) likelihood as a marginalisation over multiple layers of latent variables z. And does not provide an evaluation of the error involved in the process when the model is misspecified. As a minor supplementary appeal of the paper, the use of an asymmetric Galton quincunx to illustrate an intractable array of latent variables will certainly induce me to exploit it in projects and courses!

[Disclaimer: I was not involved in the PNAS editorial process at any point!]

BayesComp’20

First, I really have to congratulate my friend Jim Hobert for a great organisation of the meeting adopting my favourite minimalist principles (no name tag, no “goodies” apart from the conference schedule, no official talks). Without any pretense at objectivity, I also appreciated very much the range of topics and the sweet frustration of having to choose between two or three sessions each time. Here are some notes taken during some talks (with no implicit implication for the talks no mentioned, re. above frustration! as well as very short nights making sudden lapse in concentration highly likely).

On Day 1, Paul Fearnhead’s inaugural plenary talk was on continuous time Monte Carlo methods, mostly bouncy particle and zig-zag samplers, with a detailed explanation on the simulation of the switching times which likely brought the audience up to speed even if they had never heard of them. And an opening on PDMPs used as equivalents to reversible jump MCMC, reminding me of the continuous time (point process) solutions of Matthew Stephens for mixture inference (and of Preston, Ripley, Møller).

The same morn I heard of highly efficient techniques to handle very large matrices and p>n variables selections by Akihiko Nishimura and Ruth Baker on a delayed acceptance ABC, using a cheap proxy model. Somewhat different from indirect inference. I found the reliance on ESS somewhat puzzling given the intractability of the likelihood (and the low reliability of the frequency estimate) and the lack of connection with the “real” posterior. At the same ABC session, Umberto Picchini spoke on a joint work with Richard Everitt (Warwick) on linking ABC and pseudo-marginal MCMC by bootstrap. Actually, the notion of ABC likelihood was already proposed as pseudo-marginal ABC by Anthony Lee, Christophe Andrieu and Arnaud Doucet in the discussion of Fearnhead and Prangle (2012) but I wonder at the focus of being unbiased when the quantity is not the truth, i.e. the “real” likelihood. It would seem more appropriate to attempt better kernel estimates on the distribution of the summary itself. The same session also involved David Frazier who linked our work on ABC for misspecified models and an on-going investigation of synthetic likelihood.

Later, there was a surprise occurrence of the Bernoulli factory in a talk by Radu Herbei on Gaussian process priors with accept-reject algorithms, leading to exact MCMC, although the computing implementation remains uncertain. And several discussions during the poster session, incl. one on the planning of a 2021 workshop in Oaxaca centred on objective Bayes advances as we received acceptance of our proposal by BIRS today!

On Day 2, David Blei gave a plenary introduction to variational Bayes inference and latent Dirichlet allocations, somewhat too introductory for my taste although other participants enjoyed this exposition. He also mentioned a recent JASA paper on the frequentist consistency of variational Bayes that I should check. Speaking later with PhD students, they really enjoyed this opening on an area they did not know that well.

A talk by Kengo Kamatani (whom I visited last summer) on improved ergodicity rates for heavy tailed targets and Crank-NIcholson modifications to the random walk proposal (which uses an AR(1) representation instead of the random walk). With the clever idea of adding the scale of the proposal as an extra parameter with a prior of its own. Gaining one order of magnitude in the convergence speed (i.e. from d to 1 and from d² to d, where d is the dimension), which is quite impressive (and just published in JAP).Veronica Rockova linked Bayesian variable selection and machine learning via ABC, with conditions on the prior for model consistency. And a novel approach using part of the data to learn an ABC partial posterior, which reminded me of the partial  Bayes factors of the 1990’s although it is presumably unrelated. And a replacement of the original rejection ABC via multi-armed bandits, where each variable is represented by an arm, called ABC Bayesian forests. Recalling the simulation trick behind Thompson’s approach, reproduced for the inclusion or exclusion of variates and producing a fixed estimate for the (marginal) inclusion probabilities, which makes it sound like a prior-feeback form of empirical Bayes. Followed by a talk of Gregor Kastner on MCMC handling of large time series with specific priors and a massive number of parameters.

The afternoon also had a wealth of exciting talks and missed opportunities (in the other sessions!). Which ended up with a strong if unintended French bias since I listened to Christophe Andrieu, Gabriel Stolz, Umut Simsekli, and Manon Michel on different continuous time processes, with Umut linking GANs, multidimensional optimal transport, sliced-Wasserstein, generative models, and new stochastic differential equations. Manon Michel gave a highly intuitive talk on creating non-reversibility, getting rid of refreshment rates in PDMPs to kill any form of reversibility.

distilling importance

As I was about to leave Warwick at the end of last week, I noticed a new arXival by Dennis Prangle, distilling importance sampling. In connection with [our version of] population Monte Carlo, “each step of [Dennis’] distilled importance sampling method aims to reduce the Kullback Leibler (KL) divergence from the distilled density to the current tempered posterior.”  (The introduction of the paper points out various connections with ABC, conditional density estimation, adaptive importance sampling, X entropy, &tc.)

“An advantage of [distilled importance sampling] over [likelihood-free] methods is that it performs inference on the full data, without losing information by using summary statistics.”

A notion used therein I had not heard before is the one of normalising flows, apparently more common in machine learning and in particular with GANs. (The slide below is from Shakir Mohamed and Danilo Rezende.) The  notion is to represent an arbitrary variable as the bijective transform of a standard variate like a N(0,1) variable or a U(0,1) variable (calling the inverse cdf transform). The only link I can think of is perfect sampling where the representation of all simulations as a function of a white noise vector helps with coupling.

I read a blog entry by Eric Jang on the topic (who produced this slide among other things) but did not emerge much the wiser. As the text instantaneously moves from the Jacobian formula to TensorFlow code… In Dennis’ paper, it appears that the concept is appealing for quickly producing samples and providing a rich family of approximations, especially when neural networks are included as transforms. They are used to substitute for a tempered version of the posterior target, validated as importance functions and aiming at being the closest to this target in Kullback-Leibler divergence. With the importance function interpretation, unbiased estimators of the gradient [in the parameter of the normalising flow] can be derived, with potential variance reduction. What became clearer to me from reading the illustration section is that the prior x predictive joint can also be modeled this way towards producing reference tables for ABC (or GANs) much faster than with the exact model. (I came across several proposals of that kind in the past months.) However, I deem mileage should vary depending on the size and dimension of the data. I also wonder at the connection between the (final) distribution simulated by distilled importance [the least tempered target?] and the ABC equivalent.