Proceedings of ICML 2021 Workshop on Theoretic Foundation, Criticism, and Application Trend of Explainable AI

Deep neural networks lack foundational understanding: Despite their widespread success in computer vision and language processing, the basic principles explaining why DNNs work and how they maintain performance remain largely unknown, creating a significant gap in AI development.

Adversarial vulnerabilities are a critical concern: The workshop highlights that DNNs' resilience to adversarial attacks (intentional manipulations designed to fool AI systems) is poorly understood, suggesting these systems may be more fragile than their strong performance metrics indicate.

Explainability remains controversial and contested: The proceedings emphasize that interpreting how DNNs make decisions is both "compelling yet controversial," indicating the field lacks consensus on the best methods for making AI systems transparent and trustworthy.

Theoretical foundations must precede widespread deployment: The workshop identifies limitations in current explainable AI (XAI) approaches as "new bottlenecks" blocking future development, suggesting that organizations deploying AI systems may be moving faster than our ability to understand and explain them responsibly.

Multiple perspectives are needed on XAI: By bringing together work on theoretic foundations, criticisms, and application trends, the workshop recognizes that responsible AI requires input from diverse viewpoints—not just technical solutions, but also critical examination of XAI's limitations and real-world implications.