Master Class: Lever­ag­ing Opti­miza­tion Tech­niques to Scale Bayesian Inference (Emily Fox)

SpeakerEmily Fox
AffiliationUniversity of Washington
DateThursday, 02 Jul 2015
Time16:00 - 17:00
LocationRoom 1.02, Malet Place Engineering Building
Event seriesMaster Class: Carlos Guestrin (2-3 July 2015)
Description

Data streams of increas­ing com­plex­ity are being col­lected in a vari­ety of fields rang­ing from neu­ro­science, genomics, and envi­ron­men­tal mon­i­tor­ing to e-commerce based on tech­nolo­gies and infra­struc­tures pre­vi­ously unavail­able. With the advent of Markov chain Monte Carlo (MCMC) com­bined with the com­pu­ta­tional power to imple­ment such algo­rithms, deploy­ing increas­ingly expres­sive mod­els has been a focus in recent decades. Unfor­tu­nately, tra­di­tional algo­rithms for Bayesian infer­ence in these mod­els such as MCMC and vari­a­tional infer­ence do not typ­i­cally scale to the large datasets encoun­tered in prac­tice. Like­wise, these algo­rithms are not applic­a­ble to the increas­ingly com­mon sit­u­a­tion where an unbounded amount of data arrive as a stream and infer­ences need to be made on-the-fly. In this talk, we will present a series of algo­rithms— sto­chas­tic gra­di­ent Hamil­ton­ian Monte Carlo, HMM sto­chas­tic vari­a­tional infer­ence, and stream­ing Bayesian non­para­met­ric infer­ence— to address var­i­ous aspects of the chal­lenge in scal­ing Bayesian infer­ence; our algo­rithms focus on deploy­ing sto­chas­tic gra­di­ents and work­ing within an opti­miza­tion frame­work. We demon­strate our meth­ods on a vari­ety of appli­ca­tions includ­ing online movie rec­om­men­da­tions, seg­ment­ing a human chro­matin data set with 250 mil­lion obser­va­tions, and clus­ter­ing a stream of New York Times documents.

iCalendar csml_id_237.ics