NCAF: NTD-based Concept Activation Factorisation Framework for CNN Explainability

Fine-grained visual categorization (FGVC) tasks often feature deep convolutional neural networks (CNNs) that are black-box in nature and so, are not intrinsically interpretable/explainable. Instead of relying on feature-level descriptors, explanations can be made using human-understandable concepts. It becomes imperative to integrate factorization methods that automatically summarize the high-dimensional CNN representations to produce meaningful concepts. Linear factorization methods do not capture the non-linear relations in the activations from CNNs and so are limited for concept-based explainability. In this study, we rethink the invertible concept-based explainer (ICE) by proposing a non-negative Tucker Decomposition (NTD)based concept activation factorization (NCAF) framework for fine-grained categorization explainability. We introduce a new ingredient to the automatic concept extraction module — a nonnegative Tucker decomposition for improved concept discovery based on its nonlinear reconstruction/decomposition efficiencies. Our proposed method is validated quantitatively and qualitatively using fine-grained classification experiments that reveal maximum fidelity scores of 0.92 and 0.63 for the Stanford Dogs and the Apple foliar disease data sets respectively.