ExaM: Unsupervised Concept-Based Representation Learning to Better Explain Models in Vision Tasks

Interpreting deep black-box models is a complex task. Post-hoc methods for local (pixel space) and global (latent space) interpretation struggle to align learned features with concepts that are understandable by humans. Recently, interpretable-by-design methods have been proposed to address this issue by incorporating interpretability objectives during training. However, this typically results in decreased task performance, increased computational time, and reduced scalability. In contrast, object-centric methods aim to provide a more structured latent representation space, facilitating generalization through compositionality and intrinsic reasoning by learning causal variables. Despite their promising capabilities, these methods have not been successfully implemented as a basis for XAI models in real-world applications. We propose ExaM, an interpretable-by-design method that relies on object-centric representations. It offers improved capabilities for interpretation using an unsupervised concept discovery-based architecture. It explicitly provides Concept Activation Vectors and Concept Activation Maps to explain model predictions. Experimental results show that ExaM supports multiple tasks (e.g., image classification and re-identification). It sustains the performance of non-XAI methods across various datasets and backbone architectures, including real-time embedded models, with a negligible computational overhead. Our extensive analysis using quantitative metrics is validated by qualitative results, showing that ExaM outperforms state-of-the-art models for interpretability.