CONACyT COLOQUIO DE ESTADÍSTICA MACHINE LEARNING · MACHINE LEARNING 29 de octubre de 2014 /...
1
MACHINE LEARNING 29 de octubre de 2014 / Auditorio-IIMAS / 12:00 horas Circuito Escolar, Ciudad Universitaria PLÁTICAS Parallelizing MCMC for Bayesian nonparametrics Ph.D Sinead Williamson University of Texas at Austin, USA COLOQUIO DE ESTADÍSTICA CONACyT INSTITUTO DE INVESTIGACIONES EN MATEMÁTICAS APLICADAS Y EN SISTEMAS DE LA UNIVERSIDAD NACIONAL AUTÓNOMA DE MÉXICO Bayesian nonparametric models, such as those based on the Dirichlet process and the Pitman-Yor process, provide elegant and flexible alternatives to parametric models when the number of un- derlying components is unknown or growing. Unfortunately, inference in such models can be slow, and previous parallelization methods have relied on introducing approximations which can lead to inaccuracies in the posterior estimate. In this talk, I will construct auxiliary variable representations for the Dirichlet process, the Pitman-Yor process, and some hierarchical extensions, and show how these representations facilitate the development of distributed Markov chain Monte Carlo schemes that use the correct equilibrium distribution. Experimental analyses show that this approach allows scalable inference without the deterioration in estimate quality that accompanies existing meth- ods. Joint work with Avinava Dubey and Eric Xing.
Transcript of CONACyT COLOQUIO DE ESTADÍSTICA MACHINE LEARNING · MACHINE LEARNING 29 de octubre de 2014 /...
MACHINE LEARNING
29 de octubre de 2014 / Auditorio-IIMAS / 12:00 horasCircuito Escolar, Ciudad Universitaria
PLÁTICAS
Parallelizing MCMC for Bayesian nonparametricsPh.D Sinead WilliamsonUniversity of Texas at Austin, USA
COLOQUIO DE ESTADÍSTICACONACyT
INSTITUTO DE INVESTIGACIONES EN MATEMÁTICAS APLICADAS Y EN SISTEMASDE LA UNIVERSIDAD NACIONAL AUTÓNOMA DE MÉXICO
Bayesian nonparametric models, such as those based on the Dirichlet process and the Pitman-Yor process, provide elegant and �exible alternatives to parametric models when the number of un-derlying components is unknown or growing. Unfortunately, inference in such models can be slow, and previous parallelization methods have relied on introducing approximations which can lead to inaccuracies in the posterior estimate. In this talk, I will construct auxiliary variable representations for the Dirichlet process, the Pitman-Yor process, and some hierarchical extensions, and show how these representations facilitate the development of distributed Markov chain Monte Carlo schemes that use the correct equilibrium distribution. Experimental analyses show that this approach allows scalable inference without the deterioration in estimate quality that accompanies existing meth-ods. Joint work with Avinava Dubey and Eric Xing.
