Search Papers | Poster Sessions | All Posters
Poster A72 in Poster Session A - Tuesday, August 6, 2024, 4:15 – 6:15 pm, Johnson Ice Rink
A Graph Neural Network Framework to Model Human Reasoning
Quan Do1,2 (), Caroline Ahn1,2,3, Leah Bakst2,3,4, Michael Pascale2,3,4, Joseph McGuire1,2,3,4, Chantal Stern1,2,3,4, Michael Hasselmo1,2,3; 1Graduate Program for Neuroscience, Boston University, 2Center for Systems Neuroscience, Boston University, 3Department of Psychological and Brain Sciences, Boston University, 4Cognitive Neuroimaging Center, Boston University
When confronting a new challenge in an unfamiliar and puzzling situation, humans can rapidly formulate a hypothesis based on limited interactions and come up with a solution tailored to the specific problem. However, coming up with a quick solution does not guarantee a preferable outcome. It is therefore beneficial to study the representations and neural circuits underlying human reasoning, not only to inspire the development of machines that reason flexibly and quickly like humans, but also to understand how human reasoning may be impaired in certain circumstances or by certain disorders. Here we introduced a framework to discover the representations that could lead to reasoning success and failure in humans, and explored how these representations can be realized with neural circuits. We tested and validated the framework on a human dataset (n=220) collected in our lab on a modified version of the Abstraction and Reasoning Corpus. We found that our Message Passing Graph Neural Network, when taking in graphs encoding different relationships and different levels of abstraction, can reproduce human solutions. We then mapped the space of graph representations that lead to error modes in humans to observe the link between topology and functions in reasoning.
Keywords: graph neural network human reasoning inductive bias few-shot learning