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

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Cognitive Science

Background:

• Behavioral flexibility, the ability to switch tasks, is crucial for cognition.
• The neural mechanisms underlying this flexibility, especially inferring rules, are not well understood.
• Recurrent neural networks (RNNs) offer a computational model to explore these mechanisms.

Purpose of the Study:

• To investigate the neural circuit mechanisms enabling behavioral flexibility using trained RNNs.
• To model task-rule inference and switching in a computational framework.
• To elucidate the role of specific neuronal populations in flexible behavior.

Main Methods:

• Trained RNNs to perform a Wisconsin Card Sorting Test analog.
• Analyzed network modules for rule representation and sensorimotor mapping.
• Dissected trained networks, including excitatory and inhibitory neuron roles.
• Investigated the impact of silencing specific interneurons on network dynamics and performance.

Main Results:

• Emergence of rule representation via persistent activity, error monitoring, and gated sensorimotor mapping.
• Consistent circuit mechanisms across different training hyperparameters.
• Distinct dynamical trajectories for different rules in separate population activity subspaces.
• Silencing somatostatin-expressing interneurons collapsed representational subspaces and reduced performance.

Conclusions:

• A specific circuit mechanism, involving somatostatin-expressing interneurons, explains representational subspaces in flexible behavior.
• RNNs provide a valuable tool for understanding neural circuit mechanisms of cognitive flexibility.
• Persistent activity and specific interneuron types are critical for adaptive task switching and rule inference.