Thread by @jehosafet

Excited to share a preprint of our work "Learning is shaped by abrupt changes in neural engagement," advised by Aaron Batista, Steve Chase, and Byron Yu.

biorxiv.org/content/10.110…

I'm (even more?) excited to finally make my own #tweeprint! (1/n)

Internal states such as our attention and motivation involve brain-wide changes in neural activity. We know changes in these states can impact your behavior. For example, when someone surprises you:
(2/n)

If changes in internal states can impact immediate behavior, maybe they can also impact how you *learn* new behaviors. To learn, neural activity must change in particular ways. But what if internal state changes move you in the wrong way?
(3/n)

To study this question we used a brain-computer interface (BCI) learning paradigm in monkeys. Monkeys learned to modulate neural activity in motor cortex (M1) to move a computer cursor to different visual targets.
(4/n)

During these experiments, neural activity in M1 showed large trial-to-trial changes suggestive of an arousal-like process. These changes were correlated with pupil size, suggesting they reflected the monkey's internal state.
(5/n)

Here's how neural activity varied trial-to-trial for each of 8 targets (each dot is a trial, color indicates target), before monkeys started learning a new BCI mapping. The orange lines summarize this variability. We're calling them "neural engagement axes."
(6/n)

When we gave monkeys a new BCI mapping to use, something interesting happened: neural activity increased immediately along each engagement axis! (Black line connects activity on 1st trial to each target.) This happened even when doing so negatively impacted cursor movements.(7/n)

So neural activity varied trial-to-trial along neural engagement axes. And when monkeys first started learning, neural activity almost always increased abruptly along these axes. On later trials, it gradually subsided.
(8/n)

Did these changes in neural engagement help or hurt monkeys' ability to learn to move the cursor towards different targets? A BCI helps us here because we know the causal relationship between neural activity and behavior (i.e., cursor velocities).
(9/n)

We split targets into those where cursor speeds would be improved (T1) vs impaired (T2) by an increase in neural engagement. Monkeys learned both target types, but T1 targets were learned faster!
(10/n)

To conclude, M1 population activity showed large trial-to-trial changes, which we term "neural engagement." These changes interacted with learning, and helped to explain why some targets were learned more quickly than others.
(11/n)

Check out the paper for more details. And thanks to all my co-authors! Emily Oby, @MattGolub_Neuro, Lindsay Bahureksa, Patrick Sadtler, Kristin Quick, Stephen Ryu, Elizabeth Tyler-Kabara, Aaron Batista, Steve Chase, and Byron Yu.
(n/n)