Flexible gating between subspaces in a neural network model of internally guided task switching

Area of Science:

• Computational Neuroscience
• Cognitive Neuroscience

Background:

• Behavioral flexibility, the ability to switch between tasks, is vital for cognitive function.
• Understanding the neural circuits underlying this flexibility, especially when rules are inferred, is a significant challenge.

Purpose of the Study:

• To investigate the neural circuit mechanisms enabling behavioral flexibility and rule inference using trained recurrent neural networks (RNNs).
• To model a task analogous to the Wisconsin Card Sorting Test to understand emergent computational properties.

Main Methods:

• Training RNNs comprising distinct modules for rule representation and sensorimotor mapping, each with specific neuronal types.
• Analyzing the emergent properties of trained networks, including persistent activity, error monitoring, and gated mapping.
• Systematically dissecting network components and simulating the silencing of specific interneurons (somatostatin-expressing interneurons) to observe effects on network dynamics and performance.

Main Results:

• Trained RNNs spontaneously developed rule representation via persistent activity, error monitoring, and gated sensorimotor mapping.
• Distinct rules were represented in separate subspaces of the network's population activity.
• Silencing somatostatin-expressing interneurons caused these representational subspaces to collapse, leading to chance-level performance.

Conclusions:

• The study elucidates a specific circuit mechanism for behavioral flexibility and rule inference, involving somatostatin-expressing interneurons and representational subspaces.
• Recurrent neural networks provide a valuable framework for understanding complex cognitive functions and their underlying neural substrates.