From Network Dissection to Policy Dissection - Emergent Concepts in Deep Representations

Explore the inner workings of deep neural networks in this 44-minute lecture by Bolei Zhou from the University of California, Los Angeles. Discover how interpretable concepts emerge within deep representations trained for various tasks, including image classification, generation, and reinforcement learning. Gain insights into the potential applications of these discoveries, such as human-guided AI content creation and human-AI shared control for robotics. Presented at IPAM's Explainable AI for the Sciences: Towards Novel Insights Workshop, this talk delves into the series of work on uncovering emergent concepts in deep representations, offering a deeper understanding of AI systems and paving the way for meaningful human-AI interactions.

Bolei Zhou - From Network Dissection to Policy Dissection: Emergent Concepts in Deep Representations