Concept based explanation methods [1] constitute an elegant extension of saliency maps in xAI, providing not only ‘where’ important parts for a classification in the input are, but also ‘what’ the model sees there. One part of the computation consist of computing the non-negative matrix factorization (NMF), which is done on a fixed set of examples. In order to be able to adapt the obtained concepts in an online/incremental setting, the NMF conputation needs to be adapted.

There does exist work on computing online NMF, e.g., [2]. Hence, the goal of this work is to combine an incremental/online NMF computation with a concept based explanation framework, such as CRAFT [1] and evaluate it.

Literature

1. Fel, Thomas, et al. “Craft: Concept recursive activation factorization for explainability.” https://openaccess.thecvf.com/content/CVPR2023/papers/Fel_CRAFT_Concept_Recursive_Activation_FacTorization_for_Explainability_CVPR_2023_paper.pdf
2. Incremental Learning in the Non-negative Matrix Factorization https://link.springer.com/chapter/10.1007/978-3-642-03040-6_117