Atticus Geiger, Graduate Student, Department of Linguistics, Stanford University

Title: Causal Abstraction and Computational Explanations in Artificial Intelligence

Abstract: Theories of causal abstraction are a bridge between symbolic and connectionist models of computations, allowing for formally precise accounts of when a symbolic computation is implemented by a neural network. I will present on recent work where we both (1) analyze neural networks to determine whether they implement a hypothesized symbolic computation and (2) train neural networks to implement a target symbolic computation.