ICLR 2022 Highlights Advances in Graph Networks and Model Security
The tenth International Conference on Learning Representations showcases cutting-edge oral presentations spanning graph neural networks, cross-modal embeddings, and non-transferable learning for model verification.
The tenth International Conference on Learning Representations (ICLR) takes place virtually, gathering top researchers to share breakthroughs in artificial intelligence. The event features a highly selective oral presentation track that highlights the most impactful and innovative papers in the field.
Notable oral presentations explore a wide range of critical AI topics, including new perspectives on how graph neural networks surpass Weisfeiler-Lehman limitations and methods for discovering systematic errors through cross-modal embeddings. Other significant research introduces non-transferable learning techniques for model ownership verification and uses natural language to describe deep visual features.
Additional accepted works delve into the theoretical foundations of machine learning by providing tight convergence bounds for Minibatch versus Local SGD and characterizing the hidden convex optimization landscape of two-layer ReLU networks. These diverse studies collectively push the boundaries of deep learning theory, security, and practical application.