Interactive Learning from Explanations with Adaptive Guidance

Explanatory Interactive Learning (XIL) has emerged as a promising paradigm to bridge the gap between machine learning models and human understanding by integrating Explainable Artificial Intelligence (XAI) methods directly into the training process. Traditionally, XIL methods in computer vision rely on expert annotations specifying the evidence present in the input, collected before training starts and regardless of the model behaviour during training. This can be detrimental to the interactive nature of XIL and miss out on the opportunity of taking advantage of the intermediate information about the model during training. In this paper, we formalize XIL as an interactive learning paradigm to provide guidance on model explanations through a series of interactions with an expert user during training. Furthermore, we introduce an approach to approximate the evidence from sparse adaptive interactions collected as guiding points indicating where explanations were deemed irrelevant by the expert during training. We evaluate the proposed framework using a simulated interactive loop to explore interactions in an adaptive setting. Our results show that by taking advantage of the information provided by the model explanations during training, the proposed adaptive framework is able to match, or even exceed, the performance and explainability of XIL methods trained with access to the ground-truth evidence with fewer interactions.