BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Variational Bayes In Private Settings - Mijung Park\, University o
 f Amsterdam
DTSTART:20170125T150000Z
DTEND:20170125T160000Z
UID:TALK70334@talks.cam.ac.uk
CONTACT:Alessandro Davide Ialongo
DESCRIPTION:Bayesian methods are frequently used for analysing privacy-sen
 sitive datasets\, including medical records\, emails\, and educational dat
 a\, and there is a growing need for practical Bayesian inference algorithm
 s that protect the privacy of individuals' data. To this end\, we provide 
 a general framework for privacy-preserving variational Bayes (VB) for a la
 rge class of probabilistic models\, called the conjugate exponential (CE) 
 family. Our primary observation is that when models are in the CE family\,
  we can privatise the variational posterior distributions simply by pertur
 bing the expected sufficient statistics of the complete-data likelihood. F
 or widely used non-CE models with binomial likelihoods (e.g.\, logistic re
 gression)\, we exploit the Polya-Gamma data augmentation scheme to bring s
 uch models into the CE family\, such that inferences in the modified model
  resemble the original (private) variational Bayes algorithm as closely as
  possible. The iterative nature of variational Bayes presents a further ch
 allenge for privacy preservation\, as each iteration increases the amount 
 of noise needed. We overcome this challenge by combining: (1) a relaxed no
 tion of differential privacy\, called concentrated differential privacy\, 
 which provides a tight bound on the privacy cost of multiple VB iterations
  and thus significantly decreases the amount of additive noise\; and (2) t
 he privacy amplification effect of subsampling mini-batches from large-sca
 le data in stochastic learning. We empirically demonstrate the effectivene
 ss of our method in CE and non-CE models including latent Dirichlet alloca
 tion (LDA)\, Bayesian logistic regression\, and Sigmoid Belief Networks (S
 BNs)\, evaluated on real-world datasets.\n\nSpeaker Bio:\n\nMijung Park co
 mpleted her Ph.D. in the department of Electrical and Computer Engineering
  under the supervision of Prof. Jonathan Pillow (now at Princeton Universi
 ty) and Prof. Alan Bovik at The University of Texas at Austin. She worked 
 with Prof. Maneesh Sahani as a postdoc at the Gatsby computational neurosc
 ience unit at University College London. Currently\, she works with Prof. 
 Max Welling as a postdoc in the informatics institute at University of Ams
 terdam. Her research focuses on developing practical algorithms for privac
 y preserving data analysis. Previously\, she worked on a broad range of to
 pics including approximate Bayesian computation (ABC)\, probabilistic mani
 fold learning\, active learning for drug combinations and neurophysiology 
 experiments\, and Bayesian structure learning for sparse and smooth high d
 imensional parameters.
LOCATION:CBL Room BE-438\, Department of Engineering
END:VEVENT
END:VCALENDAR