Archive for scaling

scalable Metropolis-Hastings

Posted in Books, Statistics, Travel with tags , , , , , , , , , on February 12, 2019 by xi'an

Among the flury of arXived papers of last week (414!), including a fair chunk of papers submitted to ICML 2019, I spotted one entry by Cornish et al. on scalable Metropolis-Hastings, which Arnaud Doucet had mentioned to me yesterday when in Oxford. The paper builds on the delayed acceptance paper we wrote with Marco Banterlé, Clara Grazian and Anthony Lee, itself relying on a factorisation decomposition of the likelihood, combined with control variate accelerating techniques. The factorisation of both the target and the proposal allows for a (less efficient) Metropolis-Hastings acceptance ratio that is the product

\prod_{i=1}^m \alpha_i(\theta,\theta')

of individual Metropolis-Hastings acceptance ratios, but which allows for quicker rejection if one of the probabilities in the product is small, because the corresponding Bernoulli draw is zero with high probability. One advance made in Michel et al. (2017) [which I doubly missed] is that subsampling is achievable by thinning (as in PDMPs, where these authors have been quite active) through an algorithm of Shantikumar (1985) [described in Devroye’s bible]. Provided each Metropolis-Hastings probability can be lower bounded:

\alpha_i(\theta,\theta') \ge \exp\{-\psi_i \phi(\theta,\theta')\}

by a term where the transition φ does not depend on the index i in the product. The computing cost of the thinning process thus depends on the efficiency of the subsampling, namely whether or not the (Poisson) number of terms is much smaller than m, number of terms in the product. A neat trick in the current paper that extends the the Fukui-Todo procedure is to switch to the original Metropolis-Hastings when the overall lower bound is too small, recovering the geometric ergodicity of this original if it holds (Theorem 2.1). Another neat remark is that when using the naïve factorisation as the product of the n individual likelihoods, the resulting algorithm is sort of doomed as n grows, even with an optimal scaling of the proposals. To achieve scalability, the authors introduce a Taylor (i.e., Gaussian) approximation to each local target in the product and start the acceptance decomposition by using the resulting overall Gaussian approximation. Meaning that the remaining product is now made of ratios of targets over their local Taylor approximations, hence most likely close to one. And potentially lower-bounded by the remainder term in the Taylor expansion. Leading to the conclusion that, when everything goes well, meaning that the Taylor expansions can be conducted and the bounds derived for the appropriate expansion, the order of the Poisson scale is O(1/√n)..! The proposal for the Metropolis-Hastings move is actually tuned to the Gaussian approximation, appearing as a variant of the Langevin move or more exactly a discretization of an Hamiltonian move. Obviously, I cannot judge of the complexity in implementing this new scheme from just reading the paper, but this development on the split target is definitely an exciting prospect for handling huge datasets and their friends!

likelihood inflating sampling algorithm

Posted in Books, Statistics, University life with tags , , , , , , , , on May 24, 2016 by xi'an

My friends from Toronto Radu Craiu and Jeff Rosenthal have arXived a paper along with Reihaneh Entezari on MCMC scaling for large datasets, in the spirit of Scott et al.’s (2013) consensus Monte Carlo. They devised an likelihood inflated algorithm that brings a novel perspective to the problem of large datasets. This question relates to earlier approaches like consensus Monte Carlo, but also kernel and Weierstrass subsampling, already discussed on this blog, as well as current research I am conducting with my PhD student Changye Wu. The approach by Entezari et al. is somewhat similar to consensus Monte Carlo and the other solutions in that they consider an inflated (i.e., one taken to the right power) likelihood based on a subsample, with the full sample being recovered by importance sampling. Somewhat unsurprisingly this approach leads to a less dispersed estimator than consensus Monte Carlo (Theorem 1). And the paper only draws a comparison with that sub-sampling method, rather than covering other approaches to the problem, maybe because this is the most natural connection, one approach being the k-th power of the other approach.

“…we will show that [importance sampling] is unnecessary in many instances…” (p.6)

An obvious question that stems from the approach is the call for importance sampling, since the numerator of the importance sampler involves the full likelihood which is unavailable in most instances when sub-sampled MCMC is required. I may have missed the part of the paper where the above statement is discussed, but the only realistic example discussed therein is the Bayesian regression tree (BART) of Chipman et al. (1998). Which indeed constitutes a challenging if one-dimensional example, but also one that requires delicate tuning that leads to cancelling importance weights but which may prove delicate to extrapolate to other models.

a general framework for updating belief functions

Posted in Books, Statistics, University life with tags , , , , , , , , , on July 15, 2013 by xi'an

Pier Giovanni Bissiri, Chris Holmes and Stephen Walker have recently arXived the paper related to Sephen’s talk in London for Bayes 250. When I heard the talk (of which some slides are included below), my interest was aroused by the facts that (a) the approach they investigated could start from a statistics, rather than from a full model, with obvious implications for ABC, & (b) the starting point could be the dual to the prior x likelihood pair, namely the loss function. I thus read the paper with this in mind. (And rather quickly, which may mean I skipped important aspects. For instance, I did not get into Section 4 to any depth. Disclaimer: I wasn’t nor is a referee for this paper!)

The core idea is to stick to a Bayesian (hardcore?) line when missing the full model, i.e. the likelihood of the data, but wishing to infer about a well-defined parameter like the median of the observations. This parameter is model-free in that some degree of prior information is available in the form of a prior distribution. (This is thus the dual of frequentist inference: instead of a likelihood w/o a prior, they have a prior w/o a likelihood!) The approach in the paper is to define a “posterior” by using a functional type of loss function that balances fidelity to prior and fidelity to data. The prior part (of the loss) ends up with a Kullback-Leibler loss, while the data part (of the loss) is an expected loss wrt to l(THETASoEUR,x), ending up with the definition of a “posterior” that is

\exp\{ -l(\theta,x)\} \pi(\theta)

the loss thus playing the role of the log-likelihood.

I like very much the problematic developed in the paper, as I think it is connected with the real world and the complex modelling issues we face nowadays. I also like the insistence on coherence like the updating principle when switching former posterior for new prior (a point sorely missed in this book!) The distinction between M-closed M-open, and M-free scenarios is worth mentioning, if only as an entry to the Bayesian processing of pseudo-likelihood and proxy models. I am however not entirely convinced by the solution presented therein, in that it involves a rather large degree of arbitrariness. In other words, while I agree on using the loss function as a pivot for defining the pseudo-posterior, I am reluctant to put the same faith in the loss as in the log-likelihood (maybe a frequentist atavistic gene somewhere…) In particular, I think some of the choices are either hard or impossible to make and remain unprincipled (despite a call to the LP on page 7).  I also consider the M-open case as remaining unsolved as finding a convergent assessment about the pseudo-true parameter brings little information about the real parameter and the lack of fit of the superimposed model. Given my great expectations, I ended up being disappointed by the M-free case: there is no optimal choice for the substitute to the loss function that sounds very much like a pseudo-likelihood (or log thereof). (I thought the talk was more conclusive about this, I presumably missed a slide there!) Another great expectation was to read about the proper scaling of the loss function (since L and wL are difficult to separate, except for monetary losses). The authors propose a “correct” scaling based on balancing both faithfulness for a single observation, but this is not a completely tight argument (dependence on parametrisation and prior, notion of a single observation, &tc.)

The illustration section contains two examples, one of which is a full-size or at least challenging  genetic data analysis. The loss function is based on a logistic  pseudo-likelihood and it provides results where the Bayes factor is in agreement with a likelihood ratio test using Cox’ proportional hazard model. The issue about keeping the baseline function as unkown reminded me of the Robbins-Wasserman paradox Jamie discussed in Varanasi. The second example offers a nice feature of putting uncertainties onto box-plots, although I cannot trust very much the 95%  of the credibles sets. (And I do not understand why a unique loss would come to be associated with the median parameter, see p.25.)

Watch out: Tomorrow’s post contains a reply from the authors!