# NIPS 2009 — Day 3

Here are a few papers that interested me on the Wednesday of NIPS 2009.

**On Stochastic and Worst-case Models for Investing**(Elad Hazan, Satyen Kale). The authors extend Cover’s theory of Universal Portfolios to cover the “average case” of Geometric Brownian Motion, while still providing a regret bound equivalent to that of universal portfolios in the worst case. The approach seems to offer a viable compromise between classical Mean-Variance allocation and Cover’s rebalancing rules. However, there still remains the evaluation of the approach under transaction costs, which constitute a major problem of universal portfolios. [Paper]**An LP View of the M-best MAP problem**(Menachem Fromer, Amir Globerson). The authors frame the problem of obtaining the M most probable assignments in a probabilistic graphical model as solving a linear program on a polytope similar to the marginal polytope. Successive configurations are obtained by progressively adding constraints to the program, and the method present the distinct flavor of a Benders’ decomposition. This paper helps better understand the relationship between inference in graphical models and relaxations in linear programs. [Paper]**Variational Gaussian-process factor analysis for modeling spatio-temporal data**(Jaakko Luttinen, Alexander Ilin). This is an application of Gaussian processes to modeling climatic data over very long periods (decades), with irregular measurements spread across time and space. The authors introduce a very simple factorization that decouples time and space and allows for simple inference through a variational technique. [Paper]**Sharing Features among Dynamical Systems with Beta Processes**(Emily Fox, Erik Sudderth, Michael Jordan, Alan Willsky). This is a follow-up of a paper presented at last year’s NIPS by the same authors, in which they introduced the Sticky-HDP-HMM. This year, they serve the Beta-HMM, which provide a nonparametric Bayesian treatment of the case where several time series share states and emission distributions in an HMM (they do not cover the case of the switching linear dynamical system, as last year, but this is a logical extension of this year’s model). They apply this model to mocap data, where the model learns more-than-reasonable segmentations in a completely unsupervised fashion. Code is available on Emily Fox’s website. [Paper]**Bayesian Belief Polarization**(Alan Jern, Kai-min Chang, Charles Kemp). Amusing application in cognitive science: “common sense” shows, as an example of irrationality, the case where two people having a priori contradictory opinions reinforcing these opinons when shown identical evidence. The authors examine the class of Bayesian networks that can give rise to such inferences. They reach the conclusion that such inferences are common enough for many Bayesian networks with three variables: the key is to have a latent variable influencing the opinion variable, and itself influenced by observations. It is possible that some examples in behavioral finance or economics can be interpreted in a similar manner. [Paper]- FACTORIE: Probabilistic Programming via Imperatively Defined Factor Graphs (Andrew McCallum, Karl Schultz, Sameer Singh). This is a library of probabilistic graphical models (factor graphs) written in Scala, a functional language targeting the JVM. The interesting aspect is that contarily to dedicated probabilistic languages such as BLOG, Factorie allows plugging in new inference procedures very easily, which make it feasible to treat substructures for which efficient automatic inference is difficult. On a bibliographic coreference task, they claim performance 3-15x faster than Markov Logic Network, while reducing the number of errors by 20-25%. [Paper]

Advertisements

Posted on 2009/12/10, in Modeling and tagged Machine Learning, NIPS. Bookmark the permalink. Leave a comment.

## Leave a comment

## Comments 0