an even more senseless taxi-ride

I was (exceptionally) working in (and for) my garden when my daughter shouted down from her window that John Nash had just died. I thus completed my tree trimming and went to check about this sad item of news. What I read made the news even sadder as he and his wife had died in a taxi crash in New Jersey, apparently for not wearing seat-belts, a strategy you would think far from minimax… Since Nash was in Norway a few days earlier to receive the 2015 Abel Prize, it may even be that the couple was on its way home back from the airport.  A senseless death for a Beautiful Mind.

David Blei smile in Paris (seminar)

Nicolas Chopin just reminded me of a seminar given by David Blei in Paris tomorrow (at 4pm, SMILE seminar, INRIA 23 avenue d’Italie, 5th floor, orange room) on Stochastic Variational Inference and Scalable Topic Models, machine learning seminar that I will alas miss, being busy on giving mine at CMU. Here is the abstract:

Probabilistic topic modeling provides a suite of tools for analyzing
large collections of electronic documents.  With a collection as
input, topic modeling algorithms uncover its underlying themes and
decompose its documents according to those themes.  We can use topic
models to explore the thematic structure of a large collection of
documents or to solve a variety of prediction problems about text.

Topic models are based on hierarchical mixed-membership models,
statistical models where each document expresses a set of components
(called topics) with individual per-document proportions. The
computational problem is to condition on a collection of observed
documents and estimate the posterior distribution of the topics and
per-document proportions. In modern data sets, this amounts to
posterior inference with billions of latent variables.

How can we cope with such data?  In this talk I will describe
stochastic variational inference, a general algorithm for
approximating posterior distributions that are conditioned on massive
data sets.  Stochastic inference is easily applied to a large class of
hierarchical models, including time-series models, factor models, and
Bayesian nonparametric models.  I will demonstrate its application to
topic models fit with millions of articles.  Stochastic inference
opens the door to scalable Bayesian computation for modern data

…and from Rutgers

After my seminar in Princeton, I went to Rutgers University, in New Brunwick, New Jersey, to meet my friend Bill Strawderman and my former PhD student Aude Grelaud, and spent a pleasant evening with them. The next day, after a quick tour of the historical campus (great Old Dutch buildings!), I had a series of meetings with faculty members and with students, where we discussed extensions and applications of ABC. The seminar was on a tighter schedule than in Princeton, but we also managed to discuss the selection of summary statistics, while I insisted more on the (precision) gain brought by a reduction in the dimension of those summary statistics.

The schedule was tight as I had to catch a plane to Paris in New York (JFK) the same evening but taking advantage of the fairly efficient train facilities around New York, we still managed to share a quick beer at the Harvest Moon Brewery Café (I wish I had had time to get a tee-shirt from there!)… (The rest of the trip was 100% uneventful as I managed to sleep the whole flight back home!)

impressions from Princeton

This one-day trip to Princeton was quite profitable thanks to the exchanges I had with the members of the econometrics department there. In particular, I really appreciated the interactive way the Gregory Chow seminar was run and the way the audience got quickly focused on the central issue of ABC, namely running inference under limited information (provided by the summary statistics). What is most interesting (to me) is that (a) the discussants focused on the limiting normal distribution as a way to bypass (in)sufficiency and (b) they did not seem to deem the use of an insufficient summary statistics a major drawback of the method. I also took advantage of this trip to correct a restricted and misguided impression on the existence of unbiased estimators, to discuss about loss functions for set estimation and empirical likelihood, and to mention the interesting paradox of the normal mean norm (Example 4.2.8 in The Bayesian Choice) where the MLE based on the distribution of the estimator of the norm improves upon this initial estimator…

 

Princeton snapshot (2)