Search Papers | Poster Sessions | All Posters
Poster B20 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
Characterizing attractor geometry in human decision making via low-dimensional RNNs
Hua-Dong Xiong1, Li Ji-An2, Marcelo Mattar3, Robert Wilson1; 1University of Arizona, 2University of California San Diego, 3New York University
Recurrent neural networks (RNNs) have been widely utilized for modeling biological decision-making behaviors and uncovering underlying cognitive mechanisms. These networks demand less manual engineering and offer a more flexible framework compared to classical cognitive models such as reinforcement learning. However, previous studies have predominantly focused on simple decision making tasks, such as two-armed bandits with discrete rewards. Less is known about the ability of RNNs to uncover novel computational mechanisms in more complex settings, including tasks with multiple phases and continuous rewards. Here, we trained RNNs and classical cognitive models to predict choices of human subjects performing the Horizon task, which employs two phases to examine the human explore-exploit trade-off. Our RNNs substantially outperformed classical cognitive models. We then reverse-engineered these RNNs by distilling them into two-dimensional versions for each individual and analyzing the geometry of their attractors through dynamical systems analysis. We discovered that these RNNs identified a spectrum of correlated value-update rules and diverse forms of reward utilities. Our approach reveals diverse strategies employed by individuals, which traditional cognitive modeling often overlooks, thus advancing our understanding of complex decision-making dynamics underlying human exploration and exploitation.
Keywords: recurrent neural network dynamical system analysis decision-making computational modeling