BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Virtual Seminar: 'Bayesian pyramids: Identifying interpretable dee
 p structure underlying high-dimensional data’ - Professor David Dunson\,
  Professor of Statistical Science\, Duke University\, USA
DTSTART:20210331T130000Z
DTEND:20210331T140000Z
UID:TALK158506@talks.cam.ac.uk
CONTACT:Alison Quenault
DESCRIPTION:High dimensional categorical data are routinely collected in b
 iomedical and social sciences. It is of great importance to build interpre
 table models that perform dimension reduction and uncover meaningful laten
 t structures from such discrete data. Identifiability is a fundamental req
 uirement for valid modeling and inference in such scenarios\, yet is chall
 enging to address when there are complex latent structures. We propose a c
 lass of interpretable discrete latent structure models for discrete data a
 nd develop a general identifiability theory. Our theory is applicable to v
 arious types of latent structures\, ranging from a single latent variable 
 to deep layers of latent variables organized in a sparse graph (termed a B
 ayesian pyramid). The proposed identifiability conditions can ensure Bayes
 ian posterior consistency under suitable priors. As an illustration\, we c
 onsider the two-latent-layer model and propose a Bayesian shrinkage estima
 tion approach. Simulation results for this model corroborate identifiabili
 ty and estimability of the model parameters. Applications of the methodolo
 gy to DNA nucleotide sequence data uncover discrete latent features that a
 re both interpretable and highly predictive of sequence types. The propose
 d framework provides a recipe for interpretable unsupervised learning of d
 iscrete data\, and can be a useful alternative to popular machine learning
  methods.\n\nJoint work with Yuqi Gu
LOCATION:Virtual Seminar 
END:VEVENT
END:VCALENDAR