Mingzhang Yin
Columbia University
4:30 PM
127 Hayes-Healy
Machine Learning with Heterogeneous Datasets
Classical machine learning often relies on the i.i.d. assumption, but real-world data are often non-identically distributed. The heterogeneity, for example, may come from interventions, different subpopulations, and unmodeled spatial-temporal effects. I will present methodologies that leverage heterogeneous datasets to learn causal effects and fast adaptation ability, which help a predictive algorithm generalize to a new environment. Given a set of covariates, which are the causes of the outcome, and what is the strength of causality? We answer these questions by proposing a new optimization framework that explores invariance across heterogeneous environments. Next, we discuss an alternative approach that learns to generalize by fast adapting to environmental changes. In particular, we identify and define a prevalent task-level overfitting problem in meta-learning and describe a solution to mitigate the problem. We will illustrate the application of the proposed methods on simulated and real-world data.
View Poster