The 48th Seminar
Date and Time: January 11th 4:00 pm – 5:00 pm(JST)
Venue: Zoom webinar
Speaker: Liyuan Xu (Gatsby Computational Neuroscience Unit)
Title: Learning Deep Feature in Causal Inference with Unobserved Confounder
In order to build a trust-worthy machine learning model, it is essential to model the causal relationships between the action, or treatments, on the outcome. This is challenging when there exist unobserved confounders, which affect both treatments and the outcome and cause bias in the estimation. In this talk, we will introduce two causal methods to deal with this; One is instrumental variable regression, which uses the variable that affects the outcome only through the treatment. The other is proxy causal learning, which uses proxies, the structured side information for the confounders. We propose to learn neural network features in these methods by alternatively solving the regression problems, which avoids the need for sampling or density estimation. We also point out that these methods can be applied to the problems in reinforcement learning, such as offline-policy evaluations and confounded bandit problems.