In my work, I study generalization: how a learning agent can use related past experience to improve inferences in novel systems and situations.
No two moments are the same, yet as humans we are able to flexibly apply past experience in order to make good predictions and actions in novel and unfamiliar settings. This ability—generalization—is a core component of our cognition; some have even claimed a Universal Law. In my research, I develop computational models and mathematical formalizations of generalization in order to understand human inductive inferences better. In particular, I am interested in relational generalization: how to move beyond those things that look similar, and towards those that are abstractly or analogically the same.
In modern machine learning, powerful algorithms and hardware have recently been developed using deep learning. These learners, however, are only able to perform a single task well, after many training data and cycles, and generalize poorly. From an engineering point of view, this is unsatisfying: retraining algorithms for every problem, depending on the availability of a large dataset is an expensive solution to all of our problems. From a scientific point of view, I hope to use what I learn from studying humans—who generalize well, and can infer and act well after only a small amount of interaction with a new problem—to build better inductive biases into modern machine learners, and improve their performance, generalization, robustness, and efficiency.