Deep Interpretable Modelling

Abstract

The recent success of deep neural networks has exposed the problem of model transparency. The need for explainability is particularly critical in sensitive domains. In addition, regulatory frameworks for the “responsible” deployment of AI are emerging, creating legal requirements for transparent, explainable models. There are many approaches to explainability, including the distinction between top-down methods. Such as adapting existing logical models of explainability to deep learning and bottom-up methods (e.g., augmenting the “semantics-free” networks with fragments of semantic components). However, there is the challenge of how a deep network can learn multi-level representations or create explanation support on demand when requested. Here we describe our development and experiments with building interpretable deep neural networks for Natural Language Processing (NLP). We focus on learning interpretable representations to generate reliable explanations that give users a deeper understanding of the model’s behavior. These representations offer feature attribution, contrastive, and hierarchical explanations. We also show the effectiveness of our approach to model distillation and rationale extraction.