Seongmin Park

Institute of Cognitive Science Marc Jeannerod, CNRS, Lyon, France
CNRS email : spark@isc.cnrs.fr
Website : https://argmaxv.github.io/
Twitter : @SeongminAPark

Title

Structural abstraction and behavioral flexibility

Biosketch

Seongmin is a researcher at the Institute of Cognitive Science Marc Jeannerod, CNRS (UMR5229), where he focuses on human learning and decision-making. He received his Ph.D. from the Korea Advanced Institute of Science and Technology and completed his postdoctoral training at CNRS with Dr. Jean-Claude Dreher and at UC Davis with Dr. Erie Boorman.

Seongmin's research interests center on the neural computational mechanisms that underlie flexible decision-making and generalization. His work explores how the brain abstracts complex experiences into low-dimensional structural representations to enable generalization and flexible decision-making. These abilities are essential components of human intelligence but remain a challenge for many artificial intelligence systems and are often impaired in people with psychological disorders.

Seongmin's research has been published in peer-reviewed journals, including Nature Neuroscience and Neuron, and he was awarded the 2022 Chaire of Excellence Award from Lyon1 University.

Seongmin's goal is to establish a research program that develops neuro-computational models of flexible decision-making and generalization. His work aims to advance our understanding of the neural basis of human intelligence and improve artificial intelligence systems' ability to learn and generalize. Through interdisciplinary collaboration with other researchers, Seongmin aims to continue making significant contributions to the field of cognitive neuroscience.

Abstract

Generalizing past experiences to new situations is a hallmark of human intelligence, but it remains a challenge for many AI systems. One proposed mechanism for achieving this behavioral flexibility is through the construction of an internal model called "cognitive map"—a structural knowledge representation that indicates the relationships between discrete entities learned from different events. However, we have yet to fully understand how the brain constructs low-dimensional representations from everyday experiences and leverages its cognitive map to promote generalization and flexible decision-making. In this talk, I will present research shedding light on these questions from human neuroimaging and neural network modeling. My findings suggest that the brain organizes relationships between discrete entities into a graphical structure embedded in Euclidean space. Moreover, I will demonstrate how the geometry of the cognitive map interacts with changing task goals to facilitate flexible decision-making. Finally, I will provide evidence that the brain generalizes previously learned abstract knowledge structures to solve novel problems, akin to finding unexplored shortcuts during spatial navigation. By incorporating insights into the neural representation of cognitive maps into computational frameworks like reinforcement learning, my work indicates we can develop a deeper understanding of complex human cognition not fully accounted for by standard models. Uncovering the mechanisms underlying the brain's remarkable behavioral flexibility has implications for advancing both cognitive science and artificial intelligence.