Dr. Marcel Binz, Helmholtz Munich
YouTube Stream: https://www.youtube.com/watch?v=PYNCq_qa7tk
Join group to receive calendar invite: https://groups.google.com/a/modelingtalks.org/g/talks
Abstract:
What are the core components of human cognition? To approach this question, we build large-scale cognitive models that capture human behavior across hundreds of experiments. The winning model is based on a small set of interpretable principles: (1) a phoneme-level input representation, (2) a Hebbian-like episodic memory that stores prediction errors instead of raw content, (3) a context-dependent forgetting mechanism, (4) a selective output gate that controls when memory is allowed to guide behavior, and (5) hierarchical application of these principles. We find that these simple components are sufficient to outperform larger black-box models and to reproduce 73% of the effects found in human behavior.
Bio:
Dr. Marcel Binz is a staff scientist and deputy head of the Institute for Human-Centered AI at Helmholtz Munich. His work is situated at the intersection of cognitive science and machine learning, aiming to uncover the computational principles of the human mind. He has published over thirty scientific articles, including papers in Nature, PNAS, Behavioral and Brain Sciences, and Psychological Review, as well as leading machine learning venues such as NeurIPS, ICML, and ICLR. His research has been featured in the New York Times, Der Spiegel, Frankfurter Allgemeine Zeitung, and been adapted into a children’s version for the Science Journal for Kids.