Acquisition and performance of delayed-response tasks: a neural network model

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Cognitive modeling

Background:

• The delayed-matching-to-sample (DMS) task is a standard test for working memory and executive functions.
• Understanding the neural mechanisms underlying learning and performance in DMS is crucial for cognitive neuroscience.

Purpose of the Study:

• To develop a neurobiologically plausible neural network model of the DMS task.
• To investigate the role of executive control and working memory in learning and task performance.
• To explore the effects of neural damage on model performance.

Main Methods:

• A novel, uncommitted neural network architecture with executive and working layers was designed.
• The network learned the DMS task through reinforcement learning and reward-dependent synaptic plasticity.
• Simulations were used to analyze network dynamics, cell activity, and behavioral output.

Main Results:

• The model successfully learned the three stages of the DMS task, exhibiting stimulus specialization and top-down activity.
• Executive subnetworks were found to prevent spurious associations during learning and maintain performance in the mature model.
• The mature network demonstrated robustness against simulated cell damage, with graceful performance degradation.

Conclusions:

• Executive systems regulating information flow are essential for tasks like DMS.
• The model provides insights into prefrontal cortex control over visual processing areas.
• The model generates testable predictions regarding errors made by subjects learning the DMS task.