Using the same lesson across different actions, without confusing them

We want to apply what we have learned to new situations. Yet treating similar experiences too similarly can make us choose the wrong action. How does the brain balance the reuse of experience with the need to distinguish one situation from another?

A study published in Nature in August approached this question by observing the cerebral cortex and cerebellum of mice simultaneously. The mice performed two tasks: pushing a handle with their forelimbs and running along a course in a virtual environment. The movements differed, but both followed the same sequence: completing the action was followed by water one second later. Original paper ↗ Author materials ↗

As learning progressed, the mice began licking before the water arrived. Despite the different actions, they were predicting when the reward would come.

In the premotor cortex, the researchers observed population activity shared across the two tasks. In cerebellar granule cells, by contrast, the timing of individual cells' activity changed depending on the task.

But cerebellar activity was not reorganised independently from scratch for each task. Viewed across many cells, the activity retained a similar temporal structure while the representations of the two tasks became separated. The authors describe this as a “rotation.” This does not mean that cells or wiring physically rotate: it refers to a change in orientation in the coordinate space used to represent population activity. The separation was greater in mice that were more proficient at the tasks. Paper abstract ↗

This suggests one way to reuse experience: keep the shared temporal structure while separating its representation according to context. In the researchers' computational model, this approach also helped reconcile fast learning with a reduced tendency for new learning to disrupt another task. Authors’ modelling and analysis code ↗

We previously considered how memories can persist in a changing brain. The present study did not directly investigate long-term memory maintenance, but it offers another angle on that question. Our article on memory in a changing brain ↗

Does every cell have to remain active in the same way for an experience to remain usable as the same experience? Even if individual activity changes, could relationships preserved across the population—or the sequence through which activity unfolds—leave a structure that the brain can still use?

In artificial neural networks, learning a new task can also disrupt something the network could previously do. The idea suggested by this study is not simply to add a separate store for each task. It is to preserve what can be shared while separating what must not be confused.

The brain's flexibility may depend on more than its ability to rewrite information. Distinguishing what can change from what should remain available for reuse may be part of what allows experience to accumulate.