How to Train Deep Neural Networks to Think Strategically

Watch a 73-minute lecture from MIT professor Constantinos Daskalakis at the CMSA Conference on Mathematics in Science, exploring the intersection of Deep Learning and Game Theory. Delve into the challenges of training deep neural networks for strategic thinking, from mastering complex games like Go to defending against adversarial attacks and developing DNN-based models for human interaction. Examine why traditional Nash equilibria analysis falls short when dealing with non-concave utility functions in DNN parameters, and discover new approaches combining learning theory, complexity theory, game theory, and topological techniques. Learn about both the obstacles and opportunities that arise when training neural networks to think strategically, gaining insights into the future directions of Deep Learning and Game Theory integration.