Bridging Online Learning and Explainability in Image Classification

Explainability sheds light on the decision-making process of an opaque machine learning model. This helps humans to understand which set of features from an input sample has contributed to this decision/prediction by the model. Techniques like Grad-CAM (Gradient-weighted Class Activation Mapping) have been effective on static models but have not been extensively tested in online learning, where models evolve over incoming data. In this chapter, we aim to bridge this gap by experimenting with how we can preserve and adapt explainability in dynamic image classification networks by using the prominent Indian pines hyperspectral dataset. We propose a framework that performs model updates incrementally and generates an explanation that ensures transparency even when the model adapts to new distributions or concept drift. The experimental results show the trade-off between model adaptability and explainability. We show that post-hoc methods like Grad-CAM can be effective in online settings. This work contributes to the growing literature on the intersection of explainable AI (XAI) and Online Learning.