Discovering cognitive strategies with tiny recurrent neural networks.

Area of Science:

• Cognitive Neuroscience
• Computational Psychology
• Artificial Intelligence

Background:

• Understanding decision-making is key in neuroscience and psychology.
• Existing models like Bayesian inference and reinforcement learning have limitations in capturing realistic behavior.
• Current methods often require subjective adjustments.

Purpose of the Study:

• To develop a novel modeling approach using recurrent neural networks (RNNs) to discover cognitive algorithms in decision-making.
• To compare the performance of RNNs against classical cognitive models.
• To provide interpretable insights into the mechanisms of biological decision-making.

Main Methods:

• Utilized recurrent neural networks with a small number of units (1-4) to model learning and decision-making.
• Trained networks on six reward-learning tasks involving animal and human data.
• Interpreted trained networks using dynamical systems concepts for mechanistic understanding.

Main Results:

• Small RNNs outperformed classical cognitive models in predicting individual choices across tasks.
• RNN performance matched that of larger neural networks.
• The approach revealed interpretable cognitive strategies and estimated behavioral dimensionality.

Conclusions:

• Recurrent neural networks offer a powerful, interpretable method for discovering cognitive algorithms in decision-making.
• This approach provides a unified framework for comparing cognitive models and understanding neural mechanisms.
• It lays a foundation for studying both healthy and dysfunctional cognition.