A recurrent neural network framework for flexible and adaptive decision making based on sequence learning.

Area of Science:

• Computational neuroscience
• Cognitive science
• Artificial intelligence

Background:

• The brain requires understanding of sensory, action, and reward contingencies for adaptive behavior.
• Statistical inference and recurrent neural networks (RNNs) have shown success in natural language processing.
• The brain's computational principles for flexible behavior remain an active area of research.

Purpose of the Study:

• To investigate whether gated recurrent unit (GRU) networks can model the brain's solution to the contingency problem.
• To test a GRU network framework, inspired by natural language processing, using established behavioral tasks.
• To explore the potential of sequence learning models in understanding brain function.

Main Methods:

• Developed a GRU network framework based on statistical sequence learning principles.
• Trained network models to predict future events (sensory, action, reward) from past events.
• Evaluated network performance on four exemplar behavioral tasks used in empirical studies.

Main Results:

• GRU networks successfully reproduced animal and human behavior across tasks.
• Networks demonstrated generalization, Bayesian inference in novel conditions, and adaptive choice adjustments.
• Individual units within the networks encoded task variables and showed neurophysiological patterns consistent with empirical findings.

Conclusions:

• Neural network models based on statistical sequence learning may reflect the brain's computational principles for flexible and adaptive behavior.
• GRU networks offer a promising computational framework for understanding brain function related to contingency learning.
• This approach provides a valuable tool for investigating the neural basis of adaptive decision-making.