distributed evidence

Alexander Buchholz (who did his PhD at CREST with Nicolas Chopin), Daniel Ahfock, and my friend Sylvia Richardson published a great paper on the distributed computation of Bayesian evidence in Bayesian Analysis. The setting is one of distributed data from several sources with no communication between them, which relates to consensus Monte Carlo even though model choice has not been particularly studied from that perspective. The authors operate under the assumption of conditionally conjugate models, i.e., the existence of a data augmentation scheme into an exponential family so that conjugate priors can be used. For a division of the data into S blocks, the fundamental identity in the paper is

where α is the normalising constant of the sub-prior exp{log[p(θ)]/S} and the other terms are associated with this prior. Under the conditionally conjugate assumption, the integral can be approximated based on the latent variables. Most interestingly, the associated variance is directly connected with the variance of

under the joint:

“The variance of the ratio measures the quality of the product of the conditional sub-posterior as an importance sample proposal distribution.”

Assuming this variance is finite (which is likely). An approximate alternative is proposed, namely to replace the exact sub-posterior with a Normal distribution, as in consensus Monte Carlo, which should obviously require some consideration as to which parameterisation of the model produces the “most normal” (or the least abnormal!) posterior. And ensures a finite variance in the importance sampling approximation (as ensured by the strong bounds in Proposition 5). A problem shared by the bridgesampling package.

“…if the error that comes from MCMC sampling is relatively small and that the shard sizes are large enough so that the quality of the subposterior normal approximation is reasonable, our suggested approach will result in good approximations of the full data set marginal likelihood.”

The resulting approximation can also be handy in conjunction with reversible jump MCMC, in the sense that RJMCMC algorithms can be run in parallel on different chunks or shards of the entire dataset. Although the computing gain may be reduced by the need for separate approximations.

ISBA 2021 grand finale

Last day of ISBA (and ISB@CIRM), or maybe half-day, since there are only five groups of sessions we can attend in Mediterranean time.

My first session was one on priors for mixtures, with 162⁺ attendees at 5:15am! (well, at 11:15 Wien or Marseille time), Gertrud Malsiner-Walli distinguishing between priors on number of components [in the model] vs number of clusters [in the data], with a minor question of mine whether or not a “prior” is appropriate for a data-dependent quantity. And Deborah Dunkel presenting [very early in the US!] anchor models for fighting label switching, which reminded me of the talk she gave at the mixture session of JSM 2018 in Vancouver. (With extensions to consistency and mixtures of regression.) And Clara Grazian debating on objective priors for the number of components in a mixture [in the Sydney evening], using loss functions to build these. Overall it seems there were many talks on mixtures and clustering this year.

After the lunch break, when several ISB@CIRM were about to leave, we ran the Objective Bayes contributed session, which actually included several Stein-like minimaxity talks. Plus one by Théo Moins from the patio of CIRM, with ciccadas in the background. Incredibly chaired by my friend Gonzalo, who had a question at the ready for each and every speaker! And then the Savage Awards II session. Which ceremony is postponed till Montréal next year. And which nominees are uniformly impressive!!! The winner will only be announced in September, via the ISBA Bulletin. Missing the ISBA general assembly for a dinner in Cassis. And being back for the Bayesian optimisation session.

I would have expected more talks at the boundary of BS & ML (as well as COVID and epidemic decision making), the dearth of which should be a cause for concern if researchers at this boundary do not prioritise ISBA meetings over more generic meetings like NeurIPS… (An exception was George Papamakarios’ talk on variational autoencoders in the Savage Awards II session.)

Many many thanks to the group of students at UConn involved in setting most of the Whova site and running the support throughout the conference. It indeed went on very smoothly and provided a worthwhile substitute for the 100% on-site version. Actually, I both hope for the COVID pandemic (or at least the restrictions attached to it) to abate and for the hybrid structure of meetings to stay, along with the multiplication of mirror workshops. Being together is essential to the DNA of conferences, but travelling to a single location is not so desirable, for many reasons. Looking for ISBA 2022, a year from now, either in Montréal, Québec, or in one of the mirror sites!

ISBA 2021 low key

Fourth day of ISBA (and ISB@CIRM), which was a bit low key for me as I had a longer hike with my wife in the morning, including a swim in a sea as cold as the Annecy lake last month!, but nonetheless enjoyable and crystal clear, then attacked my pile of Biometrika submissions that had accumulated beyond the reasonable since last week, chased late participants who hadn’t paid yet, reviewed a paper that was due two weeks ago, chatted with participants before they left, discussed a research problem, and as a result ended attending only four sessions over the whole day. Including one about Models and Methods for Networks and Graphs, with interesting computation challenges, esp. in block models, the session in memoriam of Hélène Massam, where Gérard Letac (part of ISB@CIRM!), Jacek Wesolowski, and Reza Mohammadi, all coauthors of Hélène, made presentations on their joint advances. Hélène was born in Marseille, actually, in 1949, and even though she did not stay in France after her École Normale studies, it was a further commemoration to attend this session in her birth-place. I also found out about them working on the approximation of a ratio of normalising constants for the G-Wishart. The last session of my data was the Susie Bayarri memorial lecture, with Tamara Roderick as the lecturer. Reporting on an impressive bunch of tricks to reduce computing costs for hierarchical models with Gaussian processes.

ISBA 2021.3

Now on the third day which again started early with a 100% local j-ISBA session. (After a group run to and around Mont Puget, my first real run since 2020!!!) With a second round of talks by junior researchers from master to postdoc level. Again well-attended. A talk about Bayesian non-parametric sequential taxinomy by Alessandro Zito used the BayesANT acronym, which reminded me of the new vave group Adam and the Ants I was listening to forty years ago, in case they need a song as well as a logo! (Note that BayesANT is also used for a robot using Bayesian optimisation!) And more generally a wide variety in the themes. Thanks to the j-organisers of this 100% live session!

The next session was on PDMPs, which I helped organise, with Manon Michel speaking from Marseille, exploiting the symmetry around the gradient, which is distribution-free! Then, remotely, Kengo Kamatani, speaking from Tokyo, who expanded the high-dimensional scaling limit to the Zig-Zag sampler, exhibiting an argument against small refreshment rates, and Murray Pollock, from Newcastle, who exposed quite clearly the working principles of the Restore algorithm, including why coupling from the past was available in this setting. A well-attended session despite the early hour (in the USA).

Another session of interest for me [which I attended by myself as everyone else was at lunch in CIRM!] was the contributed C16 on variational and scalable inference that included a talk on hierarchical Monte Carlo fusion (with my friends Gareth and Murray as co-authors), Darren’s call to adopt functional programming in order to save Bayesian computing from extinction, normalising flows for modularisation, and Dennis’ adversarial solutions for Bayesian design, avoiding the computation of the evidence.

Wes Johnson’s lecture was about stories with setting prior distributions based on experts’ opinions. Which reminded me of the short paper Kaniav Kamary and myself wrote about ten years ago, in response to a paper on the topic in the American Statistician. And could not understand the discrepancy between two Bayes factors based on Normal versus Cauchy priors, until I was told they were mistakenly used repeatedly.

Rushing out of dinner, I attended both the non-parametric session (live with Marta and Antonio!) and the high-dimension computational session on Bayesian model choice (mute!). A bit of a schizophrenic moment, but allowing to get a rough picture in both areas. At once. Including an adaptive MCMC scheme for selecting models by Jim Griffin. Which could be run directly over the model space. With my ever-going wondering at the meaning of neighbour models.

ABC on brain networks

Research Gate sent me an automated email pointing out a recent paper citing some of our ABC papers. The paper is written by Timothy West et al., neuroscientists in the UK, comparing models of Parkinsonian circuit dynamics. Using SMC-ABC. One novelty is the update of the tolerance by a fixed difference, unless the acceptance rate is too low, in which case the tolerance is reinitialised to a starting value.

“(…) the proposal density P(θ|D⁰) is formed from the accepted parameters sets. We use a density approximation to the marginals and a copula for the joint (…) [i.e.] a nonparametric estimation of the marginal densities overeach parameter [and] the t-copula(…) Data are transformed to the copula scale (unit-square) using the kernel density estimator of the cumulative distribution function of each parameter and then transformed to the joint space with the t-copula.”

The construct of the proposal is quite involved, as described in the above quote. The model choice approach is standard (à la Grelaud et al.) but uses the median distance as a tolerance.

“(…) test whether the ABC estimator will: a) yield parameter estimates that are unique to the data from which they have been optimized; and b) yield consistent estimation of parameters across multiple instances (…) test the face validity of the model comparison framework (…) [and] demonstrate the scalability of the optimization and model comparison framework.”

The paper runs a fairly extensive test of the above features, concluding that “the ABC optimized posteriors are consistent across multiple initializations and that the output is determined by differences in the underlying model generating the given data.” Concerning model comparison, the authors mix the ABC Bayes factor with a post-hoc analysis of divergence to discriminate against overfitting. And mention the potential impact of the summary statistics in the conclusion section, albeit briefly, and the remark that the statistics were “sufficient to recover known parameters” is not supporting their use for model comparison. The additional criticism of sampling strategies for approximating Bayes factors is somewhat irrelevant, the main issue with ABC model choice being a change of magnitude in the evidence.

“ABC has established itself as a key tool for parameter estimation in systems biology (…) but is yet to see wide adoption in systems neuroscience. It is known that ABC will not perform well under certain conditions (Sunnåker et al., 2013). Specifically, it has been shown that the
simplest form of ABC algorithm based upon an rejection-sampling approach is inefficient in the case where the prior densities lie far from the true posterior (…) This motivates the use of neurobiologically grounded models over phenomenological models where often the ranges of potential parameter values are unknown.”