24 Oct 2025 · Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery
Continual learning and explainability tend to solve different problems, but they increasingly meet in the same projects. Continual learning asks how a model can keep learning without losing earlier knowledge. Explainable AI asks how the same model can remain inspectable while it does that. The review brings those two questions together and argues that they should be treated as part of the same design space.
The paper walks through the main families of methods, from prototype-based reasoning to rule extraction and representation analysis. That structure is useful because it shows that XAI is not limited to post-hoc explanations. It can also guide the learning process itself, help diagnose forgetting, and offer a cleaner way to compare the behavior of models trained on different task streams.
Why the review was useful to write
A review is especially helpful when a topic becomes broad enough that the terminology starts to drift. By organizing the field around the role of explanations in continual learning, we could describe which methods are meant to explain decisions, which methods are meant to shape training, and which methods do both at once. That makes it easier to see where the next research step can realistically happen.
XAI4CL as a practical follow-up
Alongside the article, we released XAI4CL on top of Avalanche so the ideas can be tested rather than only discussed. The library gives a practical starting point for experiments on sequential scenarios, baseline comparisons, and reproducible evaluation of XAI-guided methods.
View the XAI4CL library and see the corresponding citation in Publications.