The BayesComp MCMski V [or MCMskv for short] has now its official website, once again maintained by Merrill Lietchy from Drexel University, Philadelphia, and registration is even open! The call for contributed sessions is now over, while the call for posters remains open until the very end. The novelty from the previous post is that there will be a “Breaking news” [in-between the Late news sessions at JSM and the crash poster talks at machine-learning conferences] session to highlight major advances among poster submissions. And that there will be an opening talk by Steve [the Bayesian] Scott on the 4th, about the frightening prospect of MCMC death!, followed by a round-table and a welcome reception, sponsored by the Swiss Supercomputing Centre. Hence the change in dates. Which still allows for arrivals in Zürich on the January 4th [be with you].
Archive for Chamonix
Following the highly successful [authorised opinion!, from objective sources] MCMski IV, in Chamonix last year, the BayesComp section of ISBA has decided in favour of a two-year period, which means the great item of news that next year we will meet again for MCMski V [or MCMskv for short], this time on the snowy slopes of the Swiss town of Lenzerheide, south of Zürich. The committees are headed by the indefatigable Antonietta Mira and Mark Girolami. The plenary speakers have already been contacted and Steve Scott (Google), Steve Fienberg (CMU), David Dunson (Duke), Krys Latuszynski (Warwick), and Tony Lelièvre (Mines, Paris), have agreed to talk. Similarly, the nine invited sessions have been selected and will include Hamiltonian Monte Carlo, Algorithms for Intractable Problems (ABC included!), Theory of (Ultra)High-Dimensional Bayesian Computation, Bayesian NonParametrics, Bayesian Econometrics, Quasi Monte Carlo, Statistics of Deep Learning, Uncertainty Quantification in Mathematical Models, and Biostatistics. There will be afternoon tutorials, including a practical session from the Stan team, tutorials for which call is open, poster sessions, a conference dinner at which we will be entertained by the unstoppable Imposteriors. The Richard Tweedie ski race is back as well, with a pair of Blossom skis for the winner!
This morning, in the train to Dauphine (train that was even more delayed than usual!), I read a recent arXival of Brendon Brewer and Courtney Donovan. Entitled Fast Bayesian inference for exoplanet discovery in radial velocity data, the paper suggests to associate Matthew Stephens’ (2000) birth-and-death MCMC approach with nested sampling to infer about the number N of exoplanets in an exoplanetary system. The paper is somewhat sparse in its description of the suggested approach, but states that the birth-date moves involves adding a planet with parameters simulated from the prior and removing a planet at random, both being accepted under a likelihood constraint associated with nested sampling. I actually wonder if this actually is the birth-date version of Peter Green’s (1995) RJMCMC rather than the continuous time birth-and-death process version of Matthew…
“The traditional approach to inferring N also contradicts fundamental ideas in Bayesian computation. Imagine we are trying to compute the posterior distribution for a parameter a in the presence of a nuisance parameter b. This is usually solved by exploring the joint posterior for a and b, and then only looking at the generated values of a. Nobody would suggest the wasteful alternative of using a discrete grid of possible a values and doing an entire Nested Sampling run for each, to get the marginal likelihood as a function of a.”
This criticism is receivable when there is a huge number of possible values of N, even though I see no fundamental contradiction with my ideas about Bayesian computation. However, it is more debatable when there are a few possible values for N, given that the exploration of the augmented space by a RJMCMC algorithm is often very inefficient, in particular when the proposed parameters are generated from the prior. The more when nested sampling is involved and simulations are run under the likelihood constraint! In the astronomy examples given in the paper, N never exceeds 15… Furthermore, by merging all N’s together, it is unclear how the evidences associated with the various values of N can be computed. At least, those are not reported in the paper.
The paper also omits to provide the likelihood function so I do not completely understand where “label switching” occurs therein. My first impression is that this is not a mixture model. However if the observed signal (from an exoplanetary system) is the sum of N signals corresponding to N planets, this makes more sense.
Last week, Michael Gutmann, Ritabrata Dutta, Samuel Kaski, and Jukka Corander posted on arXiv the last version of the paper they had presented at MCMSki 4. As indicated by its (above) title, it suggests implementing ABC based on classification tools. Thus making it somewhat connected to our recent random forest paper.
The starting idea in the paper is that datasets generated from distributions with different parameters should be easier to classify than datasets generated from distributions with the same parameters. And that classification accuracy naturally induces a distance between datasets and between the parameters behind those datasets. We had followed some of the same track when starting using random forests, before realising that for our model choice setting, proceeding the entire ABC way once the random forest procedure had been constructed was counter-productive. Random forests are just too deadly as efficient model choice machines to try to compete with them through an ABC postprocessing. Performances are just… Not. As. Good!
A side question: I have obviously never thought about that before but why is the naïve Bayes classification rule so called?! It never sounded very Bayesian to me to (a) use the true value of the parameter and (b) average the classification performances. Interestingly, the authors (i) show identical performances of other classification methods (Fig. 2) and (ii) an exception for MA time series: when we first experimented random forests, raw data from an MA(2) model was tested to select between MA(1) and MA(2) models, and the performances of the resulting random forest were quite poor.
Now, an opposition between our two approaches is that Michael and his coauthors also include point estimation within the range of classification-based ABC inference. As we stressed in our paper, we restrict the range to classification and model choice because we do not think those machine learning tools are stable and powerful enough to perform regression and posterior probability approximation. I also see a practical weakness in the estimation scheme proposed in this new paper. Namely that the Monte Carlo gets into the way of the consistency theorem. And possibly of the simulation method itself. Another remark is that, while the authors compare the fit produced by different classification methods, there should be a way to aggregate them towards higher efficiency. Returning once more to our random forest paper, we saw improved performances each time we included a reference method, from LDA to SVMs. It would be interesting to see a (summary) variable selection version of the proposed method. A final remark is that computing time and effort do not seem to get mentioned in the paper (unless Indian jetlag confuses me more than usual). I wonder how fast the computing effort grows with the sample size to reach parametric and quadratic convergence rates.
My friends Luke Bornn, Natesh Pillai and Dawn Woodard just arXived along with Aaron Smith a short note on the convergence properties of ABC. When compared with acceptance-rejection or regular MCMC. Unsurprisingly, ABC does worse in both cases. What is central to this note is that ABC can be (re)interpreted as a pseudo-marginal method where the data comparison step acts like an unbiased estimator of the true ABC target (not of the original ABC target, mind!). From there, it is mostly an application of Christophe Andrieu’s and Matti Vihola’s results in this setup. The authors also argue that using a single pseudo-data simulation per parameter value is the optimal strategy (as compared with using several), when considering asymptotic variance. This makes sense in terms of simulating in a larger dimensional space but what of the cost of producing those pseudo-datasets against the cost of producing a new parameter? There are a few (rare) cases where the datasets are much cheaper to produce.
With Matt Moores and Kerrie Mengersen, from QUT, we wrote this short paper just in time for the MCMSki IV Special Issue of Statistics & Computing. And arXived it, as well. The global idea is to cut down on the cost of running an ABC experiment by removing the simulation of a humongous state-space vector, as in Potts and hidden Potts model, and replacing it by an approximate simulation of the 1-d sufficient (summary) statistics. In that case, we used a division of the 1-d parameter interval to simulate the distribution of the sufficient statistic for each of those parameter values and to compute the expectation and variance of the sufficient statistic. Then the conditional distribution of the sufficient statistic is approximated by a Gaussian with these two parameters. And those Gaussian approximations substitute for the true distributions within an ABC-SMC algorithm à la Del Moral, Doucet and Jasra (2012).
Across 20 125 × 125 pixels simulated images, Matt’s algorithm took an average of 21 minutes per image for between 39 and 70 SMC iterations, while resorting to pseudo-data and deriving the genuine sufficient statistic took an average of 46.5 hours for 44 to 85 SMC iterations. On a realistic Landsat image, with a total of 978,380 pixels, the precomputation of the mapping function took 50 minutes, while the total CPU time on 16 parallel threads was 10 hours 38 minutes. By comparison, it took 97 hours for 10,000 MCMC iterations on this image, with a poor effective sample size of 390 values. Regular SMC-ABC algorithms cannot handle this scale: It takes 89 hours to perform a single SMC iteration! (Note that path sampling also operates in this framework, thanks to the same precomputation: in that case it took 2.5 hours for 10⁵ iterations, with an effective sample size of 10⁴…)
Since my student’s paper on Seaman et al (2012) got promptly rejected by TAS for quoting too extensively from my post, we decided to include me as an extra author and submitted the paper to this special issue as well.
At MCMSki IV, I attended (and chaired) a session where Martyn Plummer presented some developments on cut models. As I was not sure I had gotten the idea [although this happened to be one of those few sessions where the flu had not yet completely taken over!] and as I wanted to check about a potential explanation for the lack of convergence discussed by Martyn during his talk, I decided to (re)present the talk at our “MCMSki decompression” seminar at CREST. Martyn sent me his slides and also kindly pointed out to the relevant section of the BUGS book, reproduced above. (Disclaimer: do not get me wrong here, the title is a pun on the infamous “drill, baby, drill!” and not connected in any way to Martyn’s talk or work!)
I cannot say I get the idea any clearer from this short explanation in the BUGS book, although it gives a literal meaning to the word “cut”. From this description I only understand that a cut is the removal of an edge in a probabilistic graph, however there must/may be some arbitrariness in building the wrong conditional distribution. In the Poisson-binomial case treated in Martyn’s case, I interpret the cut as simulating from
hence loosing some of the information about φ… Now, this cut version is a function of φ and θ that can be fed to a Metropolis-Hastings algorithm. Assuming we can handle the posterior on φ and the conditional on θ given φ. If we build a Gibbs sampler instead, we face a difficulty with the normalising constant m(y|φ). Said Gibbs sampler thus does not work in generating from the “cut” target. Maybe an alternative borrowing from the rather large if disparate missing constant toolbox. (In any case, we do not simulate from the original joint distribution.) The natural solution would then be to make a independent proposal on φ with target the posterior given z and then any scheme that preserves the conditional of θ given φ and y; “any” is rather wistful thinking at this stage since the only practical solution that I see is to run a Metropolis-Hasting sampler long enough to “reach” stationarity… I also remain with a lingering although not life-threatening question of whether or not the BUGS code using cut distributions provide the “right” answer or not. Here are my five slides used during the seminar (with a random walk implementation that did not diverge from the true target…):