Relational Predictive Reasoning in Brains and Machines

Explore a comprehensive lecture on relational predictive reasoning in brains and machines, delivered by Kim Stachenfeld from DeepMind and Columbia University. Delve into the multiple purposes of predictive models in deep learning, including driving representation learning and constituting mental simulators for planning and counterfactual reasoning. Examine recent research comparing representational changes in deep RL systems to those observed across brain regions during learning, with a focus on the connection between auxiliary predictive models and the hippocampus. Investigate the use of Graph Neural Networks in learning differentiable simulators for complex, large-scale dynamical systems, and discover how these models support accurate prediction, efficient optimization, and realistic simulations. Compare and contrast these predictive modeling approaches with popular deep learning techniques like transformers, gaining insights into lower-level intelligence from AI, psychology, and neuroscience perspectives.