Uncovering the neural basis of movement transitions

You're feeling thirsty. Your arm moves outwards, reaching for the cup of coffee on your desk, grasps the handle, and moves the cup smoothly back to your waiting lips.

You're walking to the bus stop, on your way to an important interview. Legs swinging, left then right, left then right, when the bus unexpectedly turns the corner ahead of you. Panicked, you break into a run, legs pumping, feet leaving the floor with each stride.

Both are transitions between movements, between the discrete movements of the arm - out, then stop, then back - and between the rhythmic movements of walking to running. But while we know much about how the brain represents and controls single movements, we know little about how it controls the transitions between them. Understanding this would help us build better intelligent prosthetics for the paralysed and disabled; and build better, more natural controllers for moving robots.

The challenge is that different movements are controlled by the same set of neurons in the brain. There are a set of neurons in your motor cortex that control arm movement. Elsewhere there are a set of neurons that create the rhythms of leg movement. Somehow, the activity of those same neurons changes from representing one movement to another, and does so smoothly, so that you do not freeze in place.

Our proposed work thus aims to tackle the intriguing problem of how a single group of neurons changes between patterns of activity so different that they each generate different movements, yet does so smoothly. To tackle this problem for discrete movements, we will study neural activity in the motor cortex of monkeys moving their arms to control a joystick. The monkey's goal is to move the joystick to hit each of four targets in a row, each movement between targets thus creating a discrete arm movement. To tackle this problem for rhythmic movements, we will study neural activity in the crawling circuit of sea-slugs escaping, changing from being still, to galloping, to crawling normally. Studying rhythmic transitions in sea-slugs has the unique advantages that we can reliably cause this escape response in the lab, and at the same time can record every output from about ten percent of all the essential neurons.

These data will let us answer some deep questions about how brains control transitions between movements. The first is to work out which pattern of neural activity creates which movement. We will develop methods to find when and how the patterns change, and compare these changes to the movements in monkeys and sea-slugs. This will reveal the basic neural "code" for transitions in movements.

The second is to understand if single neurons are important for transitions. The pattern of activity that is responsible for, say, galloping is shared among a set of neurons; and approximately the same pattern can be created by different combinations of those neurons. So it may be that only the pattern is consistently created, and not the activity of individual neurons. Knowing this will help us better understand how to decode movements from brain activity.

The third is to discover what physical changes to the circuit create the changes in activity pattern. We will use models of circuits to pick apart whether the timing and type of changes between movements are caused by changes to the inputs to the circuit, changes to the wiring between neurons, or something else. These insights this will help us design better ways to control changes between movements by controlling changes in brain activity.

By revealing how brains successfully and smoothly move bodies between movements, our results could provide a wealth of new options for the control of artificial or robotic limbs by patients, and for designing controllers for movement in robots.

Animals naturally perform movements as a continuous sequence, smoothly transitioning from one to the next. We know much about the neural correlates of isolated movements, about how the periodicity of population activity in motor centres encodes rhythmic movements, and about how cortical population activity encodes discrete movements like reaching. But we know little about the neural correlates of transitions between rhythmic or discrete movements.

In this project, we will take advantage of newly available datasets of population activity during transitions in rhythmic and discrete movements to determine how a neural population reconfigures its dynamics to support multiple end-goal movements. For rhythmic movements, we have dense single-neuron, single-spike resolution imaging data from the motor circuits of the marine mollusc Aplysia during their transitions from quiescence, to an escape gallop to exploratory crawling. For discrete movements, we have microelectrode array recordings from the motor cortex of monkeys performing a task with multiple, sequential arm movements.

Our proposed work will determine if distinctly observable movements map onto distinct dynamics of the neural population. We will use a dynamical systems framework to compare hypotheses for how transitions between discrete and rhythmic movements are encoded by the population. And we will define the contribution of single neurons to test the hypothesis that movement transitions are encoded by the joint activity of a population, and not by individual neurons. Finally, we will fit network models to the recorded population activity to infer how the underlying circuits change to create the transition in movement. Collectively, our findings will place strong constraints on theories for the neural control of movement, and on future designs of invasive brain-machine interfaces.

Economic & societal impact:

1. Brain-Computer Interfaces
A potential medium-term impact is in the design of invasive Brain Computer Interfaces. In human patients and primate test subjects, accurate control of robot arms and prosthetic devices is achieved by decoding invasively recorded activity of small neural populations, typically from cortex. These decoders assume a single, discrete behaviour. Our results will show how neural populations control the transition between behaviours, and will suggest the most relevant level of description for accessing the encoded information. They will thus likely inform the design of better decoders for Brain-Computer Interfaces, and extend them to handling smooth transitions between behaviours. Such advances would be applicable to any brain-computer interface that is dependent on neural population activity.

2. Robot controllers.
A further medium-term impact is potentially on the design of robot controllers. A broad swathe of robotics researchers have drawn on inspiration from neuroscience when designing control architectures for mobile robots for both rhythmic (e.g. walking and swimming) and discrete (e.g. reaching) movements. But these are largely limited to the control of single behaviours, with disjoint transitions between them. Our results will provide insight into how single neural circuits can control and move between multiple behaviours by transitioning between multiple different dynamical states. Moreover, they will give insight into the how these transitions are controlled within the circuit. Thus our work has the potential to provide a plethora of new options for the design of controllers for mobile robots.

3. Neural-inspired artificial intelligence (AI)
Neural-inspired AI, including deep learning and recurrent neural networks, is being harnessed by both the public and private sector to drive innovations in technology and healthcare. As one example, the UK-based DeepMind company recently demonstrated neural-inspired AI that could self-learn to outperform skilled human players on a range of video games, and to beat expert human players at Go. But a major bottleneck in advancing neural-inspired AI is understanding what even these relatively simple networks are learning, thus limiting attempts to improve performance or design to blind trial-and-error. This bottleneck has led to the idea of AI Neuroscience, of using the analysis tools of neuroscience on artificial neural networks in order to understand them. The techniques we develop here will be specifically designed to open the black box of the dynamics of large-scale networks of neurons. They will thus be equally applicable in AI Neuroscience, and long-term could potentially inform new advances in cutting-edge, commercial AI.

Researcher career development:
Our proposed work programme contains a number of specific elements for the career development of the postdoctoral researcher. The cross-disciplinary training will create a valuable skill set transferable within and outside academia, especially coding in industry standard MATLAB and Python languages, use of high-performance computing, development of machine learning techniques, and application of advanced statistics. The researcher will also gain experience in project management, event co-ordination, and the supervision of students, enhancing future fellowship and tenured position applications.