Atticus Geiger

In my research, I aim to explain artificial neural networks by uncovering interpretable symbolic algorithms they implement. I use formal models of causality as a common language for representing neural networks and symbolic algorithms and use a theory of causal abstraction to precisely understand the relationship between the two. I have both developed analysis methods for uncovering symbolic structure in networks and training algorithms for inducing such symbolic structure. My work contributes to an important thread of research that aims to make models more reliable, safe, and trustworthy.