Neuroscience
Building compositional tasks with shared neural subspaces
This study unveils how shared neural subspaces enable flexible task switching in monkeys, highlighting compositional neural representations across tasks.
Imagine your brain has a toolkit of LEGO bricks. These bricks represent small groups of brain cells that work together. To build a 'car' (one task), you combine a 'wheel' brick, an 'engine' brick, and a 'chassis' brick. To build an 'airplane' (a different task), you don't need a whole new set of parts. You can reuse the 'engine' brick, but combine it with a 'wing' brick and a 'fuselage' brick. This study found that the brain works similarly, reusing the same neural 'bricks' (called subspaces) in different combinations to handle various tasks, making it incredibly efficient and adaptable.
- 01
Neural representations of task-relevant information are shared across multiple tasks.
- 02
Monkeys can switch tasks by engaging shared neural subspaces.
- 03
Task-specific engagement of sensory and motor subspaces is guided by internal beliefs about tasks.
- 04
Neural subspaces encode both stimulus features and motor actions.
- 05
Flexible task performance is achieved through compositional neural representations.
Cognition is highly flexible—we perform many different tasks and continuously adapt to changing demands. Artificial neural networks trained to perform multiple tasks will reuse representations and computational components across tasks. By composing tasks from these subcomponents, an agent can flexibly switch between tasks and rapidly learn new tasks. This study shows that the same subspaces of neural activity represent task-relevant information across multiple tasks, engaging these subspaces in a task-specific manner.