In the plane to Montréal, today, I read this paper by Kulkarni, Saeedi and Gershman, which will be presented at AISTATS. The main idea is to create a mix between particle Monte Carlo and a kind of quasi-Monte Carlo technique (qNC is not mentionned in the paper), using variational inference (and coordinate ascent) to optimise the location and weight of the particles. It is however restricted to cases with finite support (as a product of N latent variables) as in an HMM with a finite state space. There is also something I do not get in the HMM case, which is that the variational approximation to the filtering is contracted sequentially. This means that at time the K highest weight current particles are selected while the past remains unchanged. Is this due to the Markovian nature of the hidden model? (Blame oxygen deprivation, altitude dizziness or travelling stress, then!) I also fail to understand how for filtering, “at each time step, the algorithm selects the K continuations (new variable assignments of the current particle set) that maximize the variational free energy.” Because the weight function to be optimised (eqn (11)) seems to freeze the whole past path of particles… I presume I will find an opportunity while in Reykjavik to discuss those issues with the authors.
Archive for Reykjavik
The next AISTATS conference will take place in Reykjavik, Iceland. On April 22-25, 2014. This conference “is an interdisciplinary gathering of researchers at the intersection of computer science, artificial intelligence, machine learning, statistics, and related areas.” The deadline for paper submissions is November 1, 2013. And there is a deadline for late-breaking poster abstract submissions, namely January 24. Given my heavy travel schedule next year, I am not sure I can attend, but I am definitely tempted! Esp. since I missed AISTATS 2013 in Phoenix, where I was kindly invited, due to The Accident…
An announcement for an Icelandic meeting next September, meeting I would have loved to attend (darn!)… This meeting is sponsored by the BayesComp session, of course!!!
We are pleased to announce that the University of Iceland will host the 3rd Workshop on Bayesian Inference for Latent Gaussian Models with Applications (LGM).
The workshop will be held in Reykjavik, Iceland, on September 12-14 2013 at Harpa ~V Reykjavik Concert Hall and Conference Centre:
The emphasized topics of LGM 2013 are:
-Spatial and spatio-temporal modeling
-Latent Gaussian models
-The workshop is not restricted to these topics
The invited speakers are:
-Matthias Katzfuß at Universität Heidelberg
-Bani Mallick at Texas A&M University
-Peter Müller at University of Texas
-Michèle Sebag at INRIA Saclay, CNRS
-Matthias Seeger at École Polytechnique Fédérale de Lausanne
-Christopher Wikle at University of Missouri
Early bird fee before May 21th ~@ 375
Registration fee after May 21th ~@ 440
Student fee ~@ 250
Detailed information on the scientific program, conference field trip, organizing committee, scientific committee and meeting registration is available on the conference web-site:
Last and final day of LGM 2012: it started most auspiciously with a bit of melted snow. The big news of the day however was [for me] that the next LGM was staged to take place in Reykjavik, Iceland. Indeed, Iceland is sharing top-of-the-list with Greenland and Patagonia for the countries I want to visit, one day, and I hope I will be able to attend and to take some time off to make a trek inland… Besides this personal (and therefore uninteresting!) happy event (!), the day contained a wealth of interesting talks, from Malgorzata Roos talking about using Hellinger distance to evaluate prior sensitivity (with the side impact of driving me posting a question on StackExchange about unbiased estimators of Hellinger distances!), to Aki Vehtari discussing cross-validation and other predictive entities (and concluding on the poor performances of DIC), to Luke Bornn proposing using auxiliary variables and warping to recover stationarity, to Alex Lenkoski achieving a fairly compelling advance by noticing that Gaussian graphical model selection can be handled outside the reversible jump framework, and more (including a machine-learning talk by Valeria Vitelli, now a next door neighbour at École Centrale, on the other side of Parc de Sceaux!). The conclusion talk by Gregor Gorjanc on statistical models for animal breeding genetics was more historical than methodological, however it should work as a nice refresher each time I need to make sure I understand the genetic vocabulary!
All in all, this was a wonderful workshop! New topics that I hope to study deeper, new people I was lucky to meet, most enjoyable city and running routes, hence a highly consistent choice of LGM meeting locations!, a superb sunset on the Trondheimsfjord the last evening, terrific breakfasts that border brunches, convenient travel connections (except for the last leg on the RER train from Charles de Gaulle where I got stuck for one hour)… Just great.