Doctoral colloquium - Tamara Bíla (30.3.2026)
Monday 30.3.2026 at 13:10, Lecture room I 9
Tamara Bíla:
Towards Concept-Based Interpretability in Image Classifiers and Vision-Language-Action Models
Abstract:
We present two main projects investigating concept-based interpretability in deep neural networks. In particular, the first focuses on convolutional neural network (CNN) image classifiers, while the second examines a state-of-the-art Vision-Language-Action model for robotic planning. To this end, a multifaceted framework is proposed, employing post-hoc, model-agnostic methods. Furthermore, supervised techniques and clustering approaches are combined to evaluate the separability of emergent and human-defined concepts. To explore the semantic content of emergent (discovered) concepts, visualization and dimensionality reduction tools such as Concept Localization Maps, UMAP, and PCA are employed. The analysis highlights the potential of hybrid interpretability approaches and outlines a promising path toward understanding deep multimodal networks by building a shared vocabulary between human and machine abstractions.