Kathleeen Cullen jokes that when she was studying electrical engineering at Brown University during the 1980s, she heard a rumor that neurons use electricity. That prompted her to take a course on the brain that convinced her to major in neuroscience as well as engineering. Cullen arrived at the University of Chicago—and later at McGill for a postdoc—at a time that researchers were starting to explore how neurons in the brain react to inputs from the senses when making voluntary movements. Many earlier observations were conducted by looking at the activity of the cells in stationary animals.
In the ensuing years, Cullen’s work as a professor at McGill has specialized in studying the vestibular system that allows us to maintain our balance. Cullen has retained a fasciation with the vestibular system because of its elegant simplicity. Vestibular neurons both receive sensory input and send commands to peripheral nerves to initiate movement. In recent weeks, Cullen and colleagues published a paper online in Nature Neuroscience that demonstrates how the calculations individual neurons make in a part of the brain called the cerebellum—a region directly connected to the vestibular system—can perform the simple task of making sure our bodies are positioned where we want them to be in relation to our surroundings. Here’s an interview with Cullen, edited for clarity.
Scientific American: The paper begins with an interesting observation about how the brain allows us to acquire new skills in response to changes in the external environment. That seems still to be a major question in neuroscience that people are trying to come to grips with in various way.
Kathleen Cullen: I think so. A lot of people who are working in neuroscience are trying to understand the neural mechanisms of learning. You probably don't think about it that often but even if you're working on a tennis serve and you feel you've got the movement down perfectly, your muscles are still changing. They’re fatiguing. You have to be updating your motor commands constantly in order to deal with the fact that your motor system is dynamic and changing over time. Put another way, because the biomechanical properties of the motor system constantly change over time we need to keep it calibrated.
SA: You seem to be getting down to some of the fine details of how the motor system works at a very elemental level. So it would be good to talk about how researchers approached things previously before going on to describe what you found.
KC: Using a combination of behavioral and theoretical approaches, people have speculated that when you move, your brain keeps track in real time of how you're actually moving vs. how you intended to move in order to appropriately update and adjust your movements. We know that the intrinsic delays of feedback from sensory signals are too slow. Instead, our brains need to calculate in real time any errors in order to accomplish motor learning.
Studies in which researchers experimentally perturbed voluntary movements had suggested that our brains ensure accurate motion by computing sensory prediction errors. The sensory prediction error is the difference between the sensory inflow your brain is expecting if you generate a movement vs. the actual sensory inflow it pulls in. The proposal had been that the brain can compare these two quantities very quickly to compute a sensory prediction error signal immediately during voluntary movement.
SA: What would be an example of a sensory prediction error?
KC: Imagine a gymnast doing a back flip on a balance beam. The gymnast has done this many times and has a great appreciation, an internal model of exactly the type of sensory inflow he or she should be getting from the vestibular system, which lets one know about their three-dimensional motion through space, as well as input from the proprioceptor receptors, which provide feedback from different muscles to also give a sense of body motion.
The body’s internal model compares its expectations based on having done something before to what it’s actually pulling in from its sensory receptors. When there’s a small error, that error may be because you lost your balance a little bit or it could be because you're in a slightly different motor state than you had been when initiating some motion. So you need to recalibrate. We had a paper in the journal Current Biology that preceded this one in Nature Neuroscience, which showed that the cerebellum computes sensory prediction errors to help you maintain balance.
What we've shown in this second paper in Nature Neuroscience is that when you have a persistent error —specifically a persistent unexpected sensory error—your brain actually learns and we can watch it learn trial by trial by trial. That's what you need to have happen to learn, and is why we called the paper ‘Learning to Expect the Unexpected’. Strikingly, as the brain learns to update its internal model of what to expect, we can actually watch this updating in real time by recording from single cerebellar output neurons.
SA: What is important about recording from single neurons?
KC: Researchers like me, systems neuroscientists who often also trained as engineers or physicists, like to compute these quantities that, in a mathematical sense, explain how the brain actually does something. This is all good and fine for describing a theoretical way in which the brain may be doing something. But the most important question remains—is the brain really performing the computation in this way and if so how would this look? So what we've actually found is an essential neural correlate for these little black boxes that we draw to illustrate our theories. To see that this is actually the way that neurons are performing computations is very exciting. Basically, in our work we have been able to very cleanly bridge the gap between the computational methods that have been applied to solving the problem using engineering approaches and actual reality. We discovered that the brain does indeed perform this sort of elegant math. The fact that we can see this manifested at the level of single neurons lets us know that the brain is actually using a particular algorithm.
SA: Does the brain do it the way you expected it do it?
KC: There has been some evidence that the cerebellar cortex (the surface of the cerebellum) is involved in developing this internal model of the expected sensory consequences. The cerebellar cortex is a network of cells that has a very beautiful and almost crystalline structure. Cerebellar damage does not cause paralysis, but instead produces disorders in fine movement, equilibrium, posture, and motor learning. Researchers now have good evidence that the cerebellum is involved in encoding signals consistent with what is called a forward model. This is an idea that's developed considerable support over the last decade. What people hadn’t demonstrated before this study was the actual computation of this error signal. This demonstration is important for most current models of motor learning. That's what we've demonstrated at the level of neurons that get direct cerebellar input. We took advantage of a particular pathway where the Purkinje cells in the cerebellar cortex project directly to a specific region of the deep cerebellar nuclei. The deep cerebellar nuclei are effectively located at the base of the ‘cauliflower shaped’ cerebellar cortex – and this is where we recorded from a small group of cells, within a subdivision of the cerebellar nuclei called the rostral fastigial nucleus.
The fastigial nucleus is a sphere that's maybe a millimeter in radius and contains neurons that are very interesting because they connect the cerebellar cortex to the spinal cord and are vital for postural and head movement control. One of the really interesting things we found is that we can see the response of these neurons beautifully tracking the comparison between predictive and actual sensory feedback systems during voluntary motion.
SA: How did you do the experiment?
KC: We carried out a trial-by-trial analysis of cerebellar neurons during the execution and adaptation of voluntary head movements and found that neuronal sensitivities dynamically tracked the comparison of predictive and feedback signals. (The extent that a neuron is activated by a particular input is known as its sensitivity.) When the relationship between the motor command and resultant movement was experimentally altered, neurons robustly responded to sensory input as if the movement was externally generated. Neuronal sensitivities then declined with the same time course as the concurrent behavioral learning. These findings provide direct evidence for an elegant computation requiring the comparison of predicted and actual sensory feedback to signal unexpected sensation.
SA: Were any of these findings a surprise?
KC: People have been studying behaviors like reaching which are quite complex. Because we were looking at a relatively simple sensory motor pathway that controls head motion, we were able to see this computation very clearly. You can say maybe it was expected based on computational models, but there are many models that people have built of the brain in neuroscience that cannot be directly compared to actual neuronal responses and circuits. It’s often not possible to directly correlate actual neuronal properties with the computational models. That the link we found is so explicit is to me quite exciting.
SA: How does your work relate to that of others in the field?
KC: There are currently researchers at a number of institutions, including Johns Hopkins in Baltimore and Ludwig Maximilian University in Munich, who are using computational modeling to understand deficits in patients with damage to the cerebellum. If you look at these patients, their disabilities are consistent with an inability to calculate sensory prediction errors. By Occam’s Razor, it would appear that this computation should exist but again these sorts of computation could exist and not be evident using current techniques. So it’s exciting that we can see this playing out in real time using conventional single-unit electrophysiology. I don't think people would have guessed that we would see it so clearly.
SA: What next experiments does this work suggest?
KC: It’s one thing to show the computation has occurred. It’s like a smoking gun in a way, but we’d now like to understand how the brain actually accomplishes this computation. Our research demonstrates that the cerebellum computes unexpected motion within milliseconds so that we can send an appropriate signal to the spinal cord to rapidly adjust our balance and learn new motor skills. By understanding how this computation is accomplished, we can develop better approaches for treating patients and can potentially also translate this knowledge to advance new technologies, such as improving how robots move and perform.
SA: What about potential clinical implications?
KC: If we understand the computations that the cerebellum is doing, then we have an opportunity to understand what happens in patients with loss of cerebellar function – for example, following stroke or in brain disorders such as multiple sclerosis. In addition, this knowledge can lead to the design of better, more effective, protocols for rehabilitation or even for sports training of athletes.
It’s particularly important to continue developing and improving rehabilitation programs because at the moment, it can be a bit of the Wild West. Health professionals do what they think is best for the patient, but we do not yet fully understand how to optimize training and exercises. But by being a little more systematic so that we take advantage of the brain’s own computational algorithms, we might get much better outcomes in the future.
(In the interview, Cullen emphasized that her two postdoctoral fellows, Jessica X. Brooks and Jerome Carriot, were essential collaborators in conducting the study on sensory prediction errors.)