From Adversarial Robustness to Concept Reasoning: A Journey Towards Trustworthy AI

Deep neural networks have achieved remarkable success across a wide range of domains, yet their lack of transparency and reliability remains a major obstacle to their deployment in high-stakes applications. In this talk, I will present a research journey towards trustworthy AI through robustness, explainability, and concept-based learning.

I will first show how adversarial robustness studies expose hidden vulnerabilities of modern machine learning systems and provide valuable insights into model behavior. I will then discuss recent findings on the limitations of widely used eXplainable AI (XAI) methods, highlighting the challenges of evaluating and trusting explanations. To overcome the limitations of post hoc explanations, I will present recent work on concept-based learning, including concept bottleneck models and semantic locality mechanisms that promote more meaningful and human-understandable representations.

Building upon these ideas, I will discuss emerging challenges in securing concept-based models, demonstrating that concept-level reasoning can itself be vulnerable. These observations motivate a broader research agenda aimed at developing robust and reliable concepts as an abstraction layer between statistical learning and symbolic reasoning.

Finally, I will outline a vision for trustworthy neuro-symbolic AI, where robustness, interpretability, concept learning, and formal reasoning are integrated to produce systems whose decisions can be understood, challenged, and ultimately trusted.