Learning Optimal Interventions Under Uncertainty

A fundamental goal in data analysis is learning which actions (i.e. interventions) are optimal for producing a desired outcome within specific individuals or over an entire population. While advances in reinforcement learning, Bayesian optimization, and bandit algorithms have shown great promise, the application of such sequential methods is primarily limited to digital environments where it is easy to iterate between modeling and experimentation. Although more widely applicable, learning from a fixed (observational) dataset will inherently involve substantial uncertainty due to sample-size limits, and it is undesirable to prescribe actions whose outcomes are unclear.

In this talk, we consider such settings from a Bayesian perspective and formalize the of role of uncertainty in data-driven decisions. Adopting a Gaussian process framework, we introduce a conservative definition of the optimal intervention which can be either tailored on an individual basis or globally enacted over a population. Subsequently, we extend these ideas to structured data settings via a recurrent variational autoencoder model. In both cases, gradient methods are employed to identify the best intervention and a key theme of our approach is carefully constraining this optimization to avoid regions of high outcome-uncertainty. We apply our methods to various tasks such as: inducing desired gene expression patterns, increasing the popularity of news articles, designing therapeutic antibodies, and revising natural language sentences.

Speaker Details

Jonas Mueller is a Computer Science PhD student at MIT advised by Tommi Jaakkola and David Gifford. His research interests lie in developing machine learning methods to advance both statistical science and artificial intelligence applications, and much of his work is motivated by high-impact problems in computational biology and natural language processing. Previously, Jonas studied Math and Statistics at UC Berkeley, where he was awarded the Departmental Citation. He also recently spent time as an intern at Microsoft Research NE working on bandit algorithms with Vasilis Syrgkanis and Matt Taddy.

Date:
Speakers:
Jonas Mueller
Affiliation:
MIT