Search Papers | Poster Sessions | All Posters
Poster B19 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
Recurrent neural network and cognitive models extract different information from task behavior for predicting psychiatric traits
Ming Bo Cai1,2 (), Wei Chen2, Samuel Zorowitz3, Yuko Yotsumoto2, Stephen Keeley4, Nathaniel Daw5; 1University of Miami, 2The University of Tokyo, 3National Institute of Mental Health, 4Fordham University, 5Princeton University
Computational cognitive models that explicitly formulate decision-making processes with canonical learning and decision algorithms enjoy the advantage of the explainability of the estimated parameters. However, these models are highly constrained by a limited selection of model components, and fail to explain much of the variability in human decision-making behavior. Recently, deep neural networks have been applied to better capture these nuanced patterns in human decisions, utilizing their flexibility in approximating unknown data distributions. Here, we investigated the abilities of a traditional reinforcement learning (RL) model and recurrent neural networks (RNNs) to extract information from subjects' behavior in a sequential decision-making task to predict compulsivity. We found that RNN models outperform traditional RL models and feed forward DNNs. Further, to predict a static psychiatric trait from the dynamic sequential RNN states, we propose a series of novel training approaches that integrate the hidden unit activity of the RNN across trials and demonstrate that this integration choice is crucial to achieve good predictive power. Lastly, while using RNNs to directly process sequential decision-making data outperforms traditional RL parameters in predicting psychiatric traits, these two approaches may still extract different types of information from choice behavior. To test this hypothesis, we combine our RNN approach with traditional RL parameters fit from the choice data. We find that a model which combines the unconstrained RNNs trained on raw behavioral data with RL-theory extracted parameters achieves the greatest predictive power, suggesting domain-informed RL approaches are able to extract information that standard deep learning models cannot.
Keywords: recurrent neural network reinforcement learning computational psychiatry