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.)

maximal couplings of the Metropolis-Hastings algorithm

As a sequel to their JRSS B paper, John O’Leary, Guanyang Wang, and [my friend, co-author and former student!] Pierre E. Jacob have recently posted a follow-up paper on maximal coupling for Metropolis-Hastings algorithms, where maximal is to be understood in terms of the largest possible probability for the coupled chains to be equal, according to the bound set by the coupling inequality. It made me realise that there is a heap of very recent works in this area.

A question that came up when reading the paper with our PhD students is whether or not the coupled chains stay identical after meeting once. When facing two different targets this seems inevitable and indeed Lemma 2 seems to show that no. A strong lemma that does not [need to] state what happens outside the diagonal Δ.

One of the essential tricks is to optimise several kinds of maximal coupling, incl. one for the Bernoullesque choice of moving, as given on p.3.

Algorithm 1 came as a novelty to me as it first seemed (to me!) the two chains may never meet, but this was before I read the small prints of the transition (proposal) kernel being maximally coupled with itself. While Algorithm 2 may be the earliest example of Metropolis-Hastings coupling I have seen, namely in 1999 in Crete, in connection with a talk by Laird Breyer and Gareth Roberts at a workshop of our ESSS network. As explained by the authors, this solution is not always a maximal coupling for the reason that

min(q¹.q²) min(α¹,α²) ≤ min(q¹α¹,q²α²)

(with q for the transition kernel and α for the acceptance probability). Lemma 1 is interesting in that it describes the probability to un-meet (!) as the surface between one of the move densities and the minimum of the two.

The first solution is to couple by plain Accept-Reject with the first chain being the proposed value and if rejected [i.e. not in C] to generate from the remainder or residual of the second target, in a form of completion of acceptance-rejection (accept when above rather than below, i.e. in A or A’). This can be shown to be a maximal coupling. Another coupling using reflection residuals works better but requires some spherical structure in the kernel. A further coupling on the acceptance of the Metropolis-Hastings move seems to bring an extra degree of improvement.

In the introduction, the alternatives about the acceptance probability α(·,·), e.g. Metropolis-Hastings versus Barker, are mentioned but would it make a difference to the preferred maximal coupling when using one or the other?

A further comment is that, in larger dimensions, I mean larger than one!, a Gibbsic form of coupling could be considered. In which case it would certainly decrease the coupling probability but may still speed up the overall convergence by coupling more often. See “maximality is sometimes less important than other properties of a coupling, such as the contraction behavior when a meeting does not occur.” (p.8)

As a final pun, I noted that Vaserstein is not a typo, as Leonid Vaseršteĭn is a Russian-American mathematician, currently at Penn State.

right place, wrong version

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.

the buzz about nuzz

“…expensive in these terms, as for each root, Λ(x(s),v) (at the cost of one epoch) has to be evaluated for each root finding iteration, for each node of the numerical integral“

When using the ZigZag sampler, the main (?) difficulty is in producing velocity switch as the switches are produced as interarrival times of an inhomogeneous Poisson process. When the rate of this process cannot be integrated out in an analytical manner, the only generic approach I know is in using Poisson thinning, obtained by finding an integrable upper bound on this rate, generating from this new process and subsampling. Finding the bound is however far from straightforward and may anyway result in an inefficient sampler. This new paper by Simon Cotter, Thomas House and Filippo Pagani makes several proposals to simplify this simulation, Nuzz standing for numerical ZigZag. Even better (!), their approach is based on what they call the Sellke construction, with Tom Sellke being a probabilist and statistician at Purdue University (trivia: whom I met when spending a postdoctoral year there in 1987-1988) who also wrote a fundamental paper on the opposition between Bayes factors and p-values with Jim Berger.

“We chose as a measure of algorithm performance the largest Kolmogorov-Smirnov (KS) distance between the MCMC sample and true distribution amongst all the marginal distributions.”

The practical trick is rather straightforward in that it sums up as the exponentiation of the inverse cdf method, completed with a numerical resolution of the inversion. Based on the QAGS (Quadrature Adaptive Gauss-Kronrod Singularities) integration routine. In order to save time Kingman’s superposition trick only requires one inversion rather than d, the dimension of the variable of interest. This nuzzled version of ZIgZag can furthermore be interpreted as a PDMP per se. Except that it retains a numerical error, whose impact on convergence is analysed in the paper. In terms of Wasserstein distance between the invariant measures. The paper concludes with a numerical comparison between Nuzz and random walk Metropolis-Hastings, HMC, and manifold MALA, using the number of evaluations of the likelihood as a measure of time requirement. Tuning for Nuzz is described, but not for the competition. Rather dramatically the Nuzz algorithm performs worse than this competition when counting one epoch for each likelihood computation and better when counting one epoch for each integral inversion. Which amounts to perfect inversion, unsurprisingly. As a final remark, all models are more or less Normal, with very smooth level sets, maybe not an ideal range