Building compositional tasks with shared neural subspaces.

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Neuroscience

Background:

• Cognitive flexibility allows rapid learning and task performance.
• Artificial neural networks reuse representations and components across tasks for flexibility.
• The brain's capacity for such compositional task representation is unknown.

Purpose of the Study:

• Investigate if the brain utilizes task compositionality.
• Determine if shared neural subspaces represent task-relevant information across multiple tasks.
• Examine how the brain compositionally combines neural subspaces for task performance.

Main Methods:

• Trained monkeys to switch between three compositionally related tasks.
• Recorded neural activity during task performance.
• Analyzed neural representations for shared subspaces and task-specific transformations.

Main Results:

• Identified shared subspaces of neural activity representing stimulus features and motor actions across tasks.
• Observed transformations from shared sensory to shared motor subspaces during task execution.
• Found flexible engagement of subspaces predicted by the animal's internal task belief.

Conclusions:

• The brain employs compositional strategies to flexibly perform multiple tasks.
• Task-relevant neural representations are shared and combined compositionally.
• This mechanism supports efficient learning and adaptation to new tasks.