Controlling working memory with learned instructions

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Neuroscience

Background:

• Current neural network models of cognition require significant human input for control.
• Developing autonomous cognitive control mechanisms for these models is challenging.

Purpose of the Study:

• To present GALIS, a general-purpose framework for creating autonomous cognitive control models.
• To demonstrate GALIS's capability in learning both task content and control procedures.

Main Methods:

• GALIS models consist of interacting attractor networks ('regions') inspired by primate cerebral cortex.
• Regions learn temporal sequences, exchange information, and gate each other's functions.
• A GALIS-based n-Back task model was trained on five task variations.

Main Results:

• The GALIS n-Back model autonomously controlled learning and memory item management.
• Model performance (accuracy, response time) strongly correlated with human subjects.
• The model made testable predictions about human performance under varying task demands.

Conclusions:

• GALIS provides an effective framework for developing autonomous cognitive control in neural networks.
• This approach advances neurocomputational methods for modeling cognition.
• GALIS facilitates the creation of more sophisticated and self-regulating cognitive models.