A Pittsburgh discovery about learning comes full circle

Why are some skills easier to learn than others?

Aaron Batista and his team found part of the answer to this central question in neuroscience twelve years ago. Now, a new publication has validated their findings in humans, offering promising new approaches for guiding learning and recovery.

Finding the foundational framework

Twelve years ago, Batista, professor of bioengineering at the Swanson School of Engineering, along with Carnegie Mellon University colleagues Byron Yu and Steven Chase, published a study in Nature that reshaped how neuroscientists think about learning. Using a brain-computer interface, the team found that monkeys could learn to control a computer cursor with new patterns of brain activity only if those patterns stayed within the brain's "intrinsic manifold" — the existing landscape that defines how a network of neurons tends to fire together.

“Neurons have been known to work as ensembles, but we didn't know how rigid or flexible those ensembles were,” Batista said. “It turns out they're quite rigid.”

Activity outside of the natural manifold, however, was difficult to learn, even with extensive practice. Their finding has since become a foundational framework in neuroscience, cited across studies of motor learning, brain-computer interfaces, and artificial neural networks.

“An ensemble of neurons working together can do a lot, as long as it preserves its relationships, but breaking those ensembles and building new ones is a slow, gradual process,” Batista said. “We speculated at the time that it is outside of that natural manifold where the ability to perform new skills might reside.”

In a Nature Neuroscience news and views commentary “Neural geometry guides learning,” Batista wrote about the new study from researchers at Yale and the Université de Montréal that has now extended these concepts to the human brain for the first time.

From a limitation to new possibilities

Using non-invasive brain imaging of humans rather than implanted electrodes in monkeys, the Yale team used MRI to let volunteers steer an avatar through a virtual environment using only their own brain activity, then tested whether participants could learn a new mapping between brain activity and the avatar's movement. Participants quickly adapted when the new mapping stayed within their brain's intrinsic manifold but made little progress when it required them to generate activity outside of it.

“Computational neuroscience and AI communities found our 2014 results valuable, but we’ve long been hoping that people who work with humans would also pick up on it,” Batista said. “Being able to see that these learning principles also apply to the human brain is a huge development."

In his commentary, Batista emphasizes the study authors’ conclusions that if learning is constrained by the structure of neural population activity, better understanding of that structure could eventually offer ways to guide learning more effectively. This could then set the stage for brain-based approaches guiding new learning, potentially by working directly with patients recovering from neurodegenerative conditions or stroke.

"We already have the ability to put electrodes in human brains and help people get better, for example if they have Parkinson’s disease or epilepsy,” Batista said. “Now, if someone is recovering from a stroke and can't make the hand movements they used to, we can build on what this team just found and find a way to boost recovery for them, and for people with stroke and other neurological conditions.”