Mind in motion
This story originally appeared in Yale Engineering magazine.
When formally trained dancers perform a well-learned dance move, what kind of activity is happening in their brains? And how does it differ when they’re improvising a dance?
Or, on a different scale, when two mice perform a similar task — pulling a lever, for instance — their brain activities are likewise similar. But when they sit idly by the lever, that’s when their individuality blossoms, their neurons firing in all different ways. So why is this, exactly?
These are some of the many questions Shreya Saxena, assistant professor of biomedical engineering, is trying to figure out.
Read more • Approximately 8 minutes
Grabbing your keys off the counter and putting them in your pocket seems like a simple task that we often take for granted. But try getting a robot or prosthetic device to do so with the same fluidity, and it becomes clear how remarkably complex the motions are, as well as the brain that makes them possible. For a long time, the brain seemed like such an impossibly mysterious and ineffable thing. In many ways, it still does, but technology is slowly peeling away the layers that shroud it. For instance, Saxena noted, it is becoming clear that computations in the brain are a function of groups of neurons and non-neuronal cells working together. Nonetheless, she says, there’s still a lot more to uncover.
Saxena’s interest in the human brain was sparked as an undergraduate by a lecture on how computational modeling can help decipher the brain.
“I was inspired by just how good the brain is at achieving things, and how little we know about it,” she said, adding that she wants to understand how we can “learn from the best machine in the world.”
There are so many questions, she notes. For instance, she wonders, when tennis greats Roger Federer and Novak Djokovic serve the ball, are their brains doing similar things? (Possibly). How about when Federer serves, versus someone off the street? (Probably not). Or when a wolf chases a hare — that’s a very complex task, cognitively speaking. But how about a pack of wolves chasing the same hare — how are their brains working together on this common goal? These are the kinds of mysteries that her lab has been busy working on. To that end, her many projects include building computational models of monkey’s forelimbs, working with a dance troupe to mine more insights about the brain’s role in movement, and simulating how animals cooperate with each other for rewards.
“The machines we build — we know how they work partly because we’ve built them,” she said. “With the brain, we’re not there yet. Part of what my lab is focusing on is reverse-engineering the brain. If we know how to build it, or if we can predict the outcomes during different tasks, then we should be able to understand all about it. But it’s a tall order, so we are a very large number of people across the world who are actually working on this.”
Saxena’s lab is focused on research at the intersection of neuroscience and artificial intelligence (AI). One emphasis of her work is control theory, which involves how all the many components of the brain work together to achieve a desired result through our bodies. She spends much of her time creating computer models of the brain’s workings, with applications in sensorimotor control (how your brain combines motor commands with sensory information to get your body to move), decision- making, and social behavior.
She applies AI to neuroscience in a couple of ways. One she refers to as a data-driven approach, in which she uses AI to directly model data recorded from the brain. For the other, a goal-driven approach, AI acts like the brain itself, emulating its functions and computations in specific contexts. Each of these approaches is made possible by advances in hardware technology. We can now record activity from thousands of single neurons in different cell types, as well as the activity of groups of neurons in many regions of the brain, while simultaneously recording high-resolution behavioral videos. The incredible complexity and sheer amount of this data requires the kinds of advances made in AI to make sense of it. Also, advances in computing technology have led to more efficient algorithms and increased computational resources, better allowing AI to emulate brain function. This can help researchers understand much more about how the brain works by providing digital twins of different brain regions — in turn, this knowledge can help build better AI.
For a recent study, Saxena created μSim, a computational framework that models the motor cortex, the part of the brain most responsible for movement. Other computational models built to simulate the brain’s capacity for movement have produced plenty of valuable insights but fall short when it comes to how flexible they are, due to a lack of integrating the data with an understanding of the goal of the specific brain region. For instance, a model might be able to simulate how the brain commands an arm to reach out to a lever. A human brain that learns a similar movement, though, can take that learned behavior to do other things — move at different rates, for instance, or reach in different directions. Known as “generalizable models” to neuroscientists, such actions are beyond the capabilities of most computational frameworks.
But μSim (pronounced MU-sim) has been based on the brain’s sensorimotor loop — that is, the chain of signals carrying sensory information through the brain, computing based on this information, and finally sending motor commands to the rest of the body. μSim’s biological realism is reinforce by incorporating anatomically accurate musculoskeletal models. Because of this, the framework can capture the complexity that goes into realizing movements. This allows it to analyze and predict novel limb movements, as well as the corresponding neural activity that leads to the movements. And that makes it extremely valuable for such applications as a brain-controlled prosthetic limb.
With μSim, Saxena and her research team outlined a set of experiments where an adult male rhesus macaque monkey was trained over months with juice rewards to make cycling movements with one of his forelimbs at eight different speeds. The researchers recorded the motions and the monkey’s brain activity.
They then created a computer model of the body and the brain’s motor pathways during this task, first building a detailed anatomical model of the limb, comprising about 40 different muscles and joints. They then created a recurrent neural network — a complex computer system that processes a massive amount of data to learn specific tasks — to represent the motor cortex that controls the limb. They trained the model with a machine-learning technique known as deep reinforcement learning, which can next allow them to draw insights about the role that dopamine plays in incentivizing the desired actions through rewards.
With this system, their virtual limb reproduces the actions of the monkey’s forelimb with high accuracy. And just as people do, the models don’t just learn how to achieve any one exact movement, but to navigate the space around them and generalize their movements. Since they are emulating the sensorimotor control loop in feed- back, the system can also react accordingly to unexpected disturbances.
It might not look it, but a limb going back and forth or reaching towards a target is a deceptively complex action. Saxena noted that it took her own baby about 10 months to learn goal-directed movements.
“Compared to that, the model only took a couple of hours!” Saxena said. “So we are doing pretty well, compared to human learning.”
“Part of what my lab is focusing on is reverse engineering the brain. If we know how to build it, or if we can predict the outcomes during difficult tasks, then we should be able to understand all about it.
Shreya Saxena, assistant professor of biomedical engineering
Really complex movements — like those of ballet dancers — also fascinate Saxena. What’s happening in the brain to coordinate all those motions is a tangle of questions that still baffles researchers. Saxena began tackling some of those questions last year as part of a collaboration with the Dance Theater of Harlem. Saxena and Samuel McDougle, assistant professor of psychology, worked with the dance troupe in a presentation on dance and human cognition.
“It was a conversation between the dancers, the choreographer, the researchers, and the audience,” she said. The conversations were interspersed with the dancers performing demonstrations to illustrate motor control and the cognition of movements. “Some of these were well-learned movements, versus something you’re just learning on the fly, while we discussed the psychology and the neuroscience of all of these movements. It was amazing.”
So, what are the cognitive mechanisms when a group of dancers perform a pirouette or a grand jeté in sync? That’s the next mystery for Saxena’s lab to work on.
More Details
Published Date
May 29, 2025
