Towards Understanding Human-aligned Neural Representation in the Presence of Confounding Variables

Abstract

Deep Neural Networks (DNNs) find one out of many possible solutions to a given task such as classification. This solution is more likely to pick up on spurious features and low-level statistical patterns in the train data rather than semantic features and highlevel abstractions, resulting in poor Out-of-Distribution (OOD) performance. In this project we aim to broaden the current knowledge surrounding spurious correlations as they relate to DNNs. We do this by measuring their effect on generalization under various settings, determining the existence of subnetworks in a DNN that capture the core features and examine potential mitigation strategies. Finally, we discuss alternative approaches that are reserved for future work.