Speaker: Kevin Ellis Cornell University Ithaca, NY Title: The Role of Higher-level Knowledge in Discovery Problems: Programs and Hierarchical Bayes Abstract: Effective discovery hinges on higher-level knowledge: inducted biases, constraints, and Bayesian priors. This talk considers learning and using higher-level knowledge, within the two contexts of linguistic rules and computer programs. For linguistic rules, I present a learning approach which jointly considers learning rules for many languages, and for computer programs, I present an approach called DreamCoder which learns to solve interrelated programming problems using neurosymbolic methods. In both cases, I discuss how hierarchical Bayesian methods, applied to structured programs and symbolic rules, offer a way of sharing statistical strength across problems and acquiring aspects of hierarchical knowledge that are interpretable and generalizable. Related paper: https://dl.acm.org/doi/pdf/10.1145/3453483.3454080