Muandet

Our group’s research aims at understanding the principles that enable autonomous agents to efficiently learn from past experience and interact successfully with complex environments, and to use this understanding to design new learning algorithms. Our research spans the following themes: 1) Prediction: How do we design machine learning (ML) algorithms that are robust to distribution shifts? We are interested in domain adaption (DA), domain generalization (DG), out-of-distribution (OOD) generalization, and robustness. We employ kernel methods, especially kernel mean embedding of distributions, as a mathematical arsenal to tackle these problems. 2) Causation: How do we leverage causal relationships in improving ML models, and conversely how do we use sophisticated ML methods to aid causal inference in complex environments? We are interested in unobserved confounders in causal inference, spurious correlation in machine learning, distributional treatment effects, counterfactual inference, and algorithmic decision making. Modern quasi-experimental designs such as instrumental variable (IV), proxy variables, and regression discontinuity design (RDD) offers tools to address these problems. 3) Regulation: How do we regulate the deployment of ML models in heterogeneous world to ensure generalizable, equitable, trustworthy, and democratic AI? Conversely, how do new challenges such as feedback loops, strategic manipulations, and adversarial attacks fundamentally change the ways ML models should be trained? To gain better understanding of these problems, we are adopting techniques from algorithmic game theory, mechanism design, social choice theory and other related sub-fields of economics. We also maintain an external group website at group.krikamol.org .