ABSTRACT:

As deep learning systems improve, their applicability to critical domains is hampered because of a lack of transparency. Post-hoc explanations attempt to address this concern but they provide no guarantee of faithfulness to the model’s computations. Inherently interpretable models are an alternative but such models are often considered to be too simple to perform well. In this talk we challenge this assumption by demonstrating how to create high performance inherently interpretable models. Our methods extend concept bottlenecks, a class of inherently interpretable models, by casting their creation as a generation problem for large language models. This allows us to develop search routines for finding high performing bottlenecks. We specialize this general approach to image classification, text classification, and visual question answering. In these domains, language model guided bottleneck models perform competitively to their uninterpretable counterparts and in low-data settings even sometimes outperform them.