Concept Visualization for Enhanced Explainability in Tabular Deep Learning

The EU's GDPR and the new EU AI Act impose strict requirements on AI systems, especially in high-risk domains like healthcare.

These regulations demand transparency and meaningful explanations for automated decisions to ensure adequate human oversight.

State-of-the-art eXplainable AI (XAI) methods, such as LIME and SHAP, attempt to meet these demands with low-level explanations.

However, these methods often yield inconsistent and decoupled results from high-level concepts that domain experts like physicians use in decision making.

In contrast, Deep Learning (DL) models that generate concept-based explanations during training offer more robust and consistent outcomes.

Despite their promise, concept-oriented DL models for tabular data remain underdeveloped, with the recent Tabular Concept Bottleneck Models (TabCBMs)—a family of interpretable, self-explaining DL models designed to learn and articulate high-level concept explanations for tabular data as an emerging yet unexplored solution.

CONVEYTab will advance TabCBMs by creating the first Visual Analytics (VA) framework that enhances interactivity and tests their real-world applicability.

This framework will be realized through several VA systems tailored for medical professionals, enabling them to create, refine, explain, and compare DL model concepts in real time. This will ensure AI predictions are transparent, aligned with domain expertise, and compliant with EU regulations.

CONVEYTab adopts a strong interdisciplinary approach, integrating expertise from life sciences, computer science, and cognitive science to meet the needs of the public sector and broader society.

The research will be conducted at Utrecht University's Information and Computing Sciences Department under the supervision of Prof. Dr. Alexandru C. Telea, an influential scholar with extensive experience in DL and XAI.

This project will help patients with motion disorders and intestinal parasites and position me as a leading VA for XAI researcher.