Kavli Affiliate: Marcelo Mattar
| Authors: Hua-Dong Xiong, Li Ji-An, Marcelo G Mattar and Robert C Wilson
| Summary:
Cognitive modeling provides a formal method to articulate and test hypotheses about cognitive processes. However, accurately and reliably estimating model parameters remains challenging due to common issues in behavioral science, such as limited data, measurement noise, experimental constraints, and model complexity. These challenges hinder the interpretability of parameters in both behavioral and neural data. In this study, we systematically examine how different optimization methods influence the generalizability, identifiability, robustness, and reliability of cognitive parameters in reinforcement learning models, which are widely used in psychology and neuroscience to study the cognitive mechanisms underlying value-based decision-making. Specifically, we compare the Nelder-Mead method (fminsearch), the de facto optimization approach in cognitive modeling, with a deep learning pipeline that integrates neural networks and modern optimization techniques, evaluating both across ten diverse value-based decision-making datasets involving human and animal participants performing various bandit tasks. Surprisingly, we find that although both approaches achieve comparable predictive accuracy, they produce divergent parameter estimates. Parameters estimated via the deep learning pipeline consistently demonstrate smaller gaps between training and test performance (better generalizability), increased resistance to parameter perturbations (enhanced robustness), improved recovery of ground-truth parameters in low-data regimes (stronger identifiability), and greater consistency across repeated measurements from the same individuals (better test-retest reliability). Our findings underscore the importance of systematically evaluating cognitive parameters from multiple perspectives to ensure their meaningful interpretation. Furthermore, the results highlight a critical yet underappreciated factor in cognitive modeling: the choice of optimization algorithm can systematically influence parameter estimates, even when predictive accuracy remains constant. These findings advocate for adopting deep learning-based optimization methods in cognitive modeling to improve parameter estimation.