DeepView is a tool for visualizing high-dimensional decision boundaries developed in this group. The supervised distance formulation emphasizes differences between different classes, but variation inside a class is not necessarily well preserved in DeepView. In contrast, concept based explanations [1,2] have been shown to extract relevant variation inside a class [1,3]. Following on this observation, the goal of this work is to include a distance measure based on concepts into the DeepView framework in order to improve the visualization of inside-class variation. Using such concept-based distances has been demonstrated in [3] as a promising approach for visualization.

Literature

[1] CRAFT: Concept Recursive Activation FacTorization for Explainability (https://arxiv.org/abs/2211.10154)
[2] Archetypal SAE (https://arxiv.org/abs/2502.12892)
[3] Holistic Approach for concepts (https://arxiv.org/abs/2306.07304)