Archive for asynchronous algorithms

Hastings 50 years later

Posted in Books, pictures, Statistics, University life with tags , , , , , , , , , , , , on January 9, 2020 by xi'an

What is the exact impact of the Metropolis-Hastings algorithm on the field of Bayesian statistics? and what are the new tools of the trade? What I personally find the most relevant and attractive element in a review on the topic is the current role of this algorithm, rather than its past (his)story, since many such reviews have already appeared and will likely continue to appear. What matters most imho is how much the Metropolis-Hastings algorithm signifies for the community at large, especially beyond academia. Is the availability or unavailability of software like BUGS or Stan a help or an hindrance? Was Hastings’ paper the start of the era of approximate inference or the end of exact inference? Are the algorithm intrinsic features like Markovianity a fundamental cause for an eventual extinction because of the ensuing time constraint and the lack of practical guarantees of convergence and the illusion of a fully automated version? Or are emerging solutions like unbiased MCMC and asynchronous algorithms a beacon of hope?

In their Biometrika paper, Dunson and Johndrow (2019) recently wrote a celebration of Hastings’ 1970 paper in Biometrika, where they cover adaptive Metropolis (Haario et al., 1999; Roberts and Rosenthal, 2005), the importance of gradient based versions toward universal algorithms (Roberts and Tweedie, 1995; Neal, 2003), discussing the advantages of HMC over Langevin versions. They also recall the significant step represented by Peter Green’s (1995) reversible jump algorithm for multimodal and multidimensional targets, as well as tempering (Miasojedow et al., 2013; Woodard et al., 2009). They further cover intractable likelihood cases within MCMC (rather than ABC), with the use of auxiliary variables (Friel and Pettitt, 2008; Møller et al., 2006) and pseudo-marginal MCMC (Andrieu and Roberts, 2009; Andrieu and Vihola, 2016). They naturally insist upon the need to handle huge datasets, high-dimension parameter spaces, and other scalability issues, with links to unadjusted Langevin schemes (Bardenet et al., 2014; Durmus and Moulines, 2017; Welling and Teh, 2011). Similarly, Dunson and Johndrow (2019) discuss recent developments towards parallel MCMC and non-reversible schemes such as PDMP as highly promising, with a concluding section on the challenges of automatising and robustifying much further the said procedures, if only to reach a wider range of applications. The paper is well-written and contains a wealth of directions and reflections, including those in my above introduction. Here are some mostly disconnected directions I would have liked to see covered or more covered

  1. convergence assessment today, e.g. the comparison of various approximation schemes
  2. Rao-Blackwellisation and other post-processing improvements
  3. other approximate inference tools than the pseudo-marginal MCMC
  4. importance of the parameterisation of the problem for convergence
  5. dimension issues and connection with quasi-Monte Carlo
  6. constrained spaces of measure zero, as for instance matrix distributions imposing zeros outside a diagonal band
  7. given the rise of the machine(-learners), are exploratory and intrinsically slow algorithms like MCMC doomed or can both fields feed one another? The section on optimisation could be expanded in that direction
  8. the wasteful nature of the random walk feature of MCMC algorithms, as opposed to non-reversible kernels like HMC and other PDMPs, missing from the gradient based methods section (and can we once again learn from physicists?)
  9. finer convergence issues and hence inference difficulties with complex MCMC algorithms like Gibbs samplers with incompatible conditionals
  10. use of the Hastings ratio in other algorithms like ABC or EP (in link with the section on generalised Bayes)
  11. adapting Metropolis-Hastings methods for emerging computing tools like GPUs and quantum computers

or possibly less covered, namely data augmentation put forward when it is a special case of auxiliary variables as in slice sampling and in earlier physics literature. For instance, both probit and logistic regressions do not truly require data augmentation and are more toy examples than really challenging applications. The approach of Carlin & Chib (1995) is another illustration, which has met with recent interest, despite requiring heavy calibration (just like RJMCMC). As well as a a somewhat awkward opposition between Gibbs and Hastings, in that I am not convinced that Gibbs does not remain ultimately necessary to handle high dimension problems, in the sense that the alternative solutions like Langevin, HMC, or PDMP, or…, are relying on Euclidean assumptions for the entire vector, while a direct product of Euclidean structures may prove more adequate.

automatic adaptation of MCMC algorithms

Posted in pictures, Statistics with tags , , , , , , , on March 4, 2019 by xi'an

“A typical adaptive MCMC sampler will approximately optimize performance given the kind of sampler chosen in the first place, but it will not optimize among the variety of samplers that could have been chosen.”

Last February (2018), Dao Nguyen and five co-authors arXived a paper that I missed. On a new version of adaptive MCMC that aims at selecting a wider range of proposal kernels. Still requiring a by-hand selection of this collection of kernels… Among the points addressed, beyond the theoretical guarantees that the adaptive scheme does not jeopardize convergence to the proper target, are a meta-exploration of the set of combinations of samplers and integration of the computational speed in the assessment of each sampler. Including the very difficulty of assessing mixing. One could deem the index of the proposal as an extra (cyber-)parameter to its generic parameter (like the scale in the random walk), but the discreteness of this index makes the extension more delicate than expected. And justifies the distinction between internal and external parameters. The notion of a worst-mixing dimension is quite appealing and connects with the long-term hope that one could spend the maximum fraction of the sampler runtime over the directions that are poorly mixing, while still keeping the target as should be. The adaptive scheme is illustrated on several realistic models with rather convincing gains in efficiency and time.

The convergence tools are inspired from Roberts and Rosenthal (2007), with an assumption of uniform ergodicity over all kernels considered therein which is both strong and delicate to assess in practical settings. Efficiency is rather unfortunately defined in terms of effective sample size, which is a measure of correlation or lack thereof, but which does not relate to the speed of escape from the basin of attraction of the starting point. I also wonder at the pertinence of the estimation of the effective sample size when the chain is based on different successive kernels, since the lack of correlation may be due to another kernel. Another calibration issue is the internal clock that relates to the average number of iterations required to tune properly a specific kernel, which again may be difficult to assess in a realistic situation. A last query is whether or not this scheme could be compared with an asynchronous (and valid) MCMC approach that exploits parallel capacities of the computer.

convergences of MCMC and unbiasedness

Posted in pictures, Statistics, University life with tags , , , , , , , , , on January 16, 2018 by xi'an

During his talk on unbiased MCMC in Dauphine today, Pierre Jacob provided a nice illustration of the convergence modes of MCMC algorithms. With the stationary target achieved after 100 Metropolis iterations, while the mean of the target taking much more iterations to be approximated by the empirical average. Plus a nice connection between coupling time and convergence. Convergence to the target.During Pierre’s talk, some simple questions came to mind, from developing an “impatient user version”, as in perfect sampling, in order  to stop chains that run “forever”,  to optimising parallelisation in order to avoid problems of asynchronicity. While the complexity of coupling increases with dimension and the coupling probability goes down, the average coupling time varies but an unexpected figure is that the expected cost per iteration is of 2 simulations, irrespective of the chosen kernels. Pierre also made a connection with optimal transport coupling and stressed that the maximal coupling was for the proposal and not for the target.

patterns of scalable Bayesian inference

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

Elaine Angelino, Matthew Johnson and Ryan Adams just arXived a massive survey of 118 pages on scalable Bayesian inference, which could have been entitled Bayes for Big Data, as this monograph covers state-of-the-art computational approaches to large and complex data structures. I did not read each and every line of it, but I have already recommended it to my PhD students. Some of its material unsurprisingly draws from the recent survey by Rémi Bardenet et al. (2015) I discussed a while ago. It also relates rather frequently to the somewhat parallel ICML paper of Korattikara et al. (2014). And to the firefly Monte Carlo procedure also discussed previously here.

Chapter 2 provides some standard background on computational techniques, Chapter 3 covers MCMC with data subsets, Chapter 4 gives some entries on MCMC with parallel and distributed architectures, Chapter 5 focus on variational solutions, and Chapter 6 is about open questions and challenges.

“Insisting on zero asymptotic bias from Monte Carlo estimates of expectations may leave us swamped in errors from high variance or transient bias.”

One central theme of the paper is the need for approximate solutions, MCMC being perceived as the exact solution. (Somewhat wrongly in the sense that the product of an MCMC is at best an empirical version of the true posterior, hence endowed with a residual and incompressible variation for a given computing budget.) While Chapter 3 stresses the issue of assessing the distance to the true posterior, it does not dwell at all on computing times and budget, which is arguably a much harder problem. Chapter 4 seems to be more aware of this issue since arguing that “a way to use parallel computing resources is to run multiple sequential MCMC algorithms at once [but that this] does not reduce the transient bias in MCMC estimates of posterior expectations” (p.54). The alternatives are to use either prefetching (which was the central theme of Elaine Angelino’s thesis), asynchronous Gibbs with the new to me (?) Hogwild Gibbs algorithms (connected in Terenin et al.’s recent paper, not quoted in the paper), some versions of consensus Monte Carlo covered in earlier posts, the missing links being in my humble opinion an assessment of the worth of those solutions (in the spirit of “here’s the solution, what was the problem again?”) and once again the computing time issue. Chapter 5 briefly discusses some recent developments in variational mean field approximations, which is farther from my interests and (limited) competence, but which appears as a particular class of approximate models and thus could (and should?) relate to likelihood-free methods. Chapter 6 about the current challenges of the field is presumably the most interesting in this monograph in that it produces open questions and suggests directions for future research. For instance, opposing the long term MCMC error with the short term transient part. Or the issue of comparing different implementations in a practical and timely perspective.

optimal importance sampling

Posted in Books, Statistics, Travel, University life with tags , , , , , , on January 13, 2016 by xi'an

somewhere near Zürich, Jan. 4, 2016An arXiv file that sat for quite a while in my to-read pile is Variance reduction in SGD by distributed importance sampling by Alain et al. I had to wait for the flight to Zürich and MCMskv to get a look at it. The part of the paper that is of primary interest to me is the generalisation of the optimal importance function result

q⁰(x)∞f(x)|h(x)|

to higher dimensions. Namely, what is the best importance function for approximating the expectation of h(X) when h is multidimensional? There does exist an optimal solution when the score function is the trace of the variance matrix. Where the solution is proportional to the target density times the norm of the target integrand

q⁰(x)∞f(x)||h(x)||

The application of the result to neural networks and stochastic gradients using minibatches of the training set somehow escapes me, even though the asynchronous aspects remind me of the recent asynchronous Gibbs sampler of Terenin, Draper, and Simpson.

While the optimality obtained in the paper is mathematically clear, I am a wee bit surprised at the approach: the lack of normalising constant in the optimum means using a reweighted approximation that drifts away from the optimal score. Furthermore, this optimum is sub-optimal when compared with the component wise optimum which produces a variance of zero (if we assume the normalising constant to be available). Obviously, using the component-wise optima requires to run as many simulations as there are components in the integrand, but since cost does not seem to be central to this study…