BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY: The Grammar Variational Autoencoder &amp\; Counterfactual Fairnes
 s - Dr Matt Kusner
DTSTART:20170912T100000Z
DTEND:20170912T110000Z
UID:TALK79861@talks.cam.ac.uk
CONTACT:54031
DESCRIPTION:In this talk I'll be covering two research directions I'm real
 ly excited about. The first is on improving deep generative models for dis
 crete data using grammars\, and the second is on using causality to ensure
  that machine learning predictions aren't discriminatory. \n	In the first 
 half of the talk I will describe how generative modeling of discrete data 
 such as arithmetic expressions and molecular structures still poses signif
 icant challenges. Crucially\, state-of-the-art methods often produce outpu
 ts that are not valid. We make the key observation that frequently\, discr
 ete data can be represented as a parse tree from a context-free grammar. W
 e propose a variational autoencoder which directly encodes from and decode
 s to these parse trees\, ensuring the generated outputs are always syntact
 ically valid. Surprisingly\, we show that not only does our model more oft
 en generate valid outputs\, it also learns a more coherent latent space in
  which nearby points decode to similar discrete outputs. We demonstrate th
 e effectiveness of our learned models by showing their improved performanc
 e in Bayesian optimization for symbolic regression and molecule generation
 .\n	In the second half of the talk\, I will detail how machine learning is
  now being used in settings where previous decisions have been made that a
 re unfairly biased against certain subpopulations\, for example those of a
  particular race\, gender\, or sexual orientation. Since this past data ma
 y be biased\, machine learning predictors must account for this to avoid p
 erpetuating or creating discriminatory practices. In this paper\, we devel
 op a framework for modeling fairness using tools from causal inference. Ou
 r definition of counterfactual fairness captures the intuition that a deci
 sion is fair towards an individual if it the same in (a) the actual world 
 and (b) a counterfactual world where the individual belonged to a differen
 t demographic group. We demonstrate our framework on a real-world problem 
 of fair prediction of success in law school and on identifying discriminat
 ion in stop-and-frisk data.\n
LOCATION:CBL Room BE-438\, Department of Engineering
END:VEVENT
END:VCALENDAR