Another great day of talks and discussions at BIRS! Continuing on the themes of the workshop between delving into the further validation of those approximation techniques and the devising of ever more approximate solutions for ever more complex problems. Among the points that came clearer to me through discussion, a realisation that the synthetic likelihood perspective is not that far away from our assumptions in the consistency paper. And that a logistic version of the approach can be constructed as well. A notion I had not met before (or have forgotten I had met) is the one of early rejection ABC, which should actually be investigated more thoroughly as it should bring considerable improvement in computing time (with the caveats of calibrating the acceptance step before producing the learning sample and of characterising the output). Both Jukka Corander and Ewan Cameron reminded us of the case of models that take minutes or hours to produce one single dataset. (In his talk on some challenging applications, Jukka Corander chose to move from socks to boots!) And Jean-Michel Marin produced an illuminating if sobering experiment on the lack of proper Bayesian coverage by ABC solutions. (It appears that Ewan’s video includes a long empty moment when we went out for the traditional group photo, missing the end of his talk.)
Archive for the pictures Category
The exact title of the paper by Jovana Metrovic, Dino Sejdinovic, and Yee Whye Teh is DR-ABC: Approximate Bayesian Computation with Kernel-Based Distribution Regression. It appeared last year in the proceedings of ICML. The idea is to build ABC summaries by way of reproducing kernel Hilbert spaces (RKHS). Regressing such embeddings to the “optimal” choice of summary statistics by kernel ridge regression. With a possibility to derive summary statistics for quantities of interest rather than for the entire parameter vector. The use of RKHS reminds me of Arthur Gretton’s approach to ABC, although I see no mention made of that work in the current paper.
In the RKHS pseudo-linear formulation, the prediction of a parameter value given a sample attached to this value looks like a ridge estimator in classical linear estimation. (I thus wonder at why one would stop at the ridge stage instead of getting the full Bayes treatment!) Things get a bit more involved in the case of parameters (and observations) of interest, as the modelling requires two RKHS, because of the conditioning on the nuisance observations. Or rather three RHKS. Since those involve a maximum mean discrepancy between probability distributions, which define in turn a sort of intrinsic norm, I also wonder at a Wasserstein version of this approach.
What I find hard to understand in the paper is how a large-dimension large-size sample can be managed by such methods with no visible loss of information and no explosion of the computing budget. The authors mention Fourier features, which never rings a bell for me, but I wonder how this operates in a general setting, i.e., outside the iid case. The examples do not seem to go into enough details for me to understand how this massive dimension reduction operates (and they remain at a moderate level in terms of numbers of parameters). I was hoping Jovana Mitrovic could present her work here at the 17w5025 workshop but she sadly could not make it to Banff for lack of funding!
The ABC workshop I co-organised has now started and, despite a few last minutes cancellations, we have gathered a great crowd of researchers on the validation and expansion of ABC methods. Or ABC’ory to keep up with my naming of workshops. The videos of the talks should come up progressively on the BIRS webpage. When I did not forget to launch the recording. The program is quite open and with this size of workshop allows for talks and discussions to last longer than planned: the first days contain several expository talks on ABC convergence, auxiliary or synthetic models, summary constructions, challenging applications, dynamic models, and model assessment. Plus prepared discussions on those topics that hopefully involve several workshop participants. We had also set some time for snap-talks, to induce everyone to give a quick presentation of one’s on-going research and open problems. The first day was rather full but saw a lot of interactions and discussions during and around the talks, a mood I hope will last till Friday! Today in replacement of Richard Everitt who alas got sick just before the workshop, we are conducting a discussion on dimensional issues, part of which is made of parts of the following slides (mostly recycled from earlier talks, including the mini-course in Les Diablerets):
A news report last weekend on Nature webpage about the new science super-campus south of Paris connected with my impressions of the whole endeavour: the annual report from the Court of Auditors estimated that the 5 billion euros invested in this construct were not exactly a clever use of public [French taxpayer] money! This notion to bring a large number of [State] engineer and scientific schools from downtown Paris to the plateau of Saclay, about 25km south-west of Paris, around École Polytechnique, had some appeal, since these were and are prestigious institutions, most with highly selective entry exams, and with similar training programs, now that they have almost completely lost the specialisation that justified their parallel existences! And since a genuine university, Paris 11 Orsay, stood nearby at the bottom of the plateau. Plus, a host of startups and research branches of companies. Hence the concept of a French MIT.
However, as so often the case in Jacobin France, the move has been decided and supported by the State “top-down” rather than by the original institutions themselves. Including a big push by Nicolas Sarkozy in 2010. While the campus can be reached by public transportation like RER, the appeal of living and working on the campus is obviously less appealing to both students and staff than in a listed building in the centre of Paris. Especially when lodging and living infrastructures are yet to be completed. But the main issue is that the fragmentation of those schools, labs and institutes, in terms of leadership, recruiting, research, and leadership, has not been solved by the move, each entity remaining strongly attached to its identity, degree, networks, &tc., and definitely unwilling to merge into a super-university with a more efficient organisation of teaching and research. Which means the overall structure as such is close to invisible at the international level. This is the point raised by the State auditors. And perceived by the State which threatens to cut funding at this late stage!
This is not the only example within French higher educations institutions since most have been forced to merged into incomprehensible super-units under the same financial threat. Like Paris-Dauphine being now part of the PSL (Paris Sciences et Lettres) heterogeneous conglomerate. (I suspect one of the primary reasons for this push by central authorities was to create larger entities towards moving up in the international university rankings, which is absurd for many reasons, from the limited worth of such rankings, to the lag between the creation of a new entity and the appearance on an international university ranking, to the difficulty in ranking researchers from such institutions: in Paris-Dauphine, the address to put on papers is more than a line long, with half a dozen acronyms!)