[cs-talks] Upcoming Seminars: IVC (Tues)
fgreen1 at bu.edu
Mon Nov 23 10:57:59 EST 2015
Learning and Inference in Distance Dependent Models
Soumya Ghosh, Disney Research Boston
Tuesday, November 24, 2015 at 2pm in MCS 148
ABSTRACT: In this talk, I will discuss some of our recent work on statistical methods for unsupervised extraction of part-based representations from natural images, three–dimensional (3D) representations of articulated objects and motion capture (MoCap) sequences. Apart from strong spatio-temporal correlations, such data typically exhibit wide variability in complexity with some containing only a few constituents and others depicting complex structure. Effective models must automatically adapt to this complexity while simultaneously modeling spatio-temporal interactions.
Recent advances in Bayesian nonparametrics (BNP) provide an attractive path towards satisfying the above desiderata. In particular, I will explore the distance dependent Chinese restaurant process (ddCRP), a distribution over partitions that captures dependencies between data instances via (user-specified) affinity functions. Although exact inference in models employing the ddCRP is intractable, simple and efficient MCMC samplers are available allowing for effective posterior inference. However, learning affinity functions in distance dependent models can be challenging. I will describe our recent work that exploits advances in approximate Bayesian computation (ABC) to effectively learn such models. I will also provide examples where statistical models endowed with ddCRP priors (and their hierarchical extensions) produce state-of-the-art results on the problems of articulated 3D object segmentations and activity discovery from MoCap sequences.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the cs-talks