Figures 4 and 5 suggest LPFC encodes animal's belief and also representations are compressed based on that belief. Encoding of the belief begins during fixation period. FEF and PAR showed suppression but not aIT. We think one of the reasons for suppression is limited capacity for cognitive control.

Similar to above question, we still don't have a solid framework that compares tasks and task components across studies. Possible unsupervised (e.g. 10.1016/j.tics.2008.02.009) and supervised methods have been proposed (e.g. 10.1038/s41592-024-02318-2) .

@coganlab.bsky.social Important question that we are actively thinking about (writing a review on it). How the brain/representations create the boundary of task components is an open question. Statistical learning? Self-supervise learning? Among possible ways that brain use to create that boundary.

Culmination of a long-term project in collaboration with wonderful team of @floramb.bsky.social @timbuschman.bsky.social @adelardalan @Nikolamatmarkov @MotoakiUchimura @nathanieldaw.bsky.social @marcelomattar.bsky.social

Monkeys had to discover the task in effect, updating their internal belief based on feedback. As they learned which task was in effect, the task-relevant shared subspaces were gradually engaged and task-irrelevant information was compressed.

When performing a task, information was dynamically transformed from the relevant shared category subspace into the appropriate motor subspace. Suggests prefrontal cortex is a ‘global workspace’, where information flexibly moves between subspaces to perform different tasks.

Neural recordings found the stimulus’ color and shape category, and the motor action, were represented in separate subspaces of neural activity. These subspaces were shared across tasks – one could ‘build’ a task from the subspaces of other tasks.

So, we trained animals to perform three tasks. Each task required categorizing a stimulus input, based on either its color or shape, and then indicating the category by making one of two different types of motor responses. Tasks shared categorization and response components.

In particular, we wanted to test the hypothesis that the brain can reuse simple task ‘components’ to compositionally build complex tasks. For example, once we learn to tell if a piece of fruit is ripe, then we can use this as a component of foraging, cooking, and eating.

Humans and animals are remarkably good at performing many different tasks. On any given morning, one might transition from driving to work, to making coffee, to checking email, etc. We wanted to understand how the brain can learn and flexibly switch between multiple tasks.

Building compositional tasks with shared neural subspaces Nature - The brain can flexibly perform multiple tasks by compositionally combining task-relevant neural representations.

Thrilled that my paper is out in the @nature.com. We explored how the brain builds complex tasks by compositionally combining simpler sub-task representations. The brain flexibly performs multiple tasks by dynamically reusing neural subspaces for sensory inputs and motor actions

rdcu.be/eRVUk