Resolving the size and nature of neocortical population codes

Cortex is the source of our most basic and most advanced brain functions, of how we hear, see, and touch; of how we think, plan, and act. All arise from the combined activity of millions or billions of individual neurons. Within these gargantuan numbers, small sets of neurons have specific roles. One set might fire to a high pitched tone; one might fire to the brush of cloth on the tip of an index finger; yet another to start moving your right elbow. Our proposal asks the simple question: to do a task, how many sets, with how many roles, does the brain use?

Imagine the part of the brain necessary for doing a particular task is an orchestra playing in a sound proofed room. Our question is the same as asking: how can we work out what score they are playing? And work out the roles of each set of instruments within that score? Up till now, our brain recording technology has been like blindly lowering microphones at random next to one or two players in the orchestra, listening for a few minutes, then trying to reconstruct the entire score. Done this way, we have no idea of which type of individual instruments are involved, let alone how they interact, or group into their wood, string and other ensembles. We don't even know how big the orchestra is. So to solve the problem of reconstructing the brain's score for a task, we need to be able to record the whole orchestra of neurons at once, one microphone per neuron. We can then work out from that cacophony what ensembles and instruments they represent, their roles, and how they combine to create the full score.

Recent technological advances means that we now have the right kind of one-microphone-per-neuron data. This has been made possible by the wonderfully neat correspondence between the whiskers on a mouse's face and the way a whisker is represented in the brain. Mice can learn to find which of two spouts contains water by touching a pole with a single whisker. This single whisker is represented in their cortex by a barrel-shaped column of neurons. It is small enough that a lab has now recorded the activity of every neuron in its top half while the mice tried to get their water. As the only representation of that single whisker, it must contain all the information the mice need to solve the task. So we know these data must contain within them the brain's orchestra for this task. Our goal is to use this data to answer our question: how many sets of neurons, with how many roles, does the brain need to solve this task?

To do so, we will use so-called "unsupervised" methods, algorithms that can determine for themselves how many different sets of neurons there are in the data, how large they are, and which neurons belong to which sets. They do this by working out which neurons are consistently active at the same time. Having found the sets, we can then find out the their roles by comparing their activity with the mouse's behaviour: for example, we can work out if some sets are active while it moves its whiskers, or while it licks the water.

If we answer this question, what do we learn? We will learn about the basic building blocks of how cortex computes. If we can only represent N things in N sets of neurons, then that places an upper limit on our capacity to think. We will learn about the resilience of cortex to damage, whether through accidents or diseases such as dementia. If multiple sets of neurons have the same task, then we may lose some and carry on as normal. But if some sets have a unique role, then damage to them, however small, could be disastrous. Ultimately, we will learn about how these sets combine to produce the full score. Labs and clinics are exploring how we can transmit the activity of small bits of motor cortex to give patients direct control over their artificial limbs. If we knew how to work out the full score for controlling limb movement, the accuracy of this control would improve many times over.

Human cortex has 20 billion neurons, but cortical neurons are typically studied one, or a handful, at a time. The building blocks of cortical function must reside between these extremes - in populations of neurons. Basic questions about cortical populations are unanswered. How many are engaged by a particular task? Are the computational roles they play distinct or overlapping? And how do they emerge over learning? Our aim is to answer these basic questions, using data from Karel Svoboda's lab that exhaustively sampled neuron activity from mice performing a tactile task.

Mice learnt to determine the location of liquid reward by detecting the position of a pole with a single spared whisker. Simultaneous multi-plane, multi-photon imaging of somatosensory cortex captured volumes containing the entire layer 2/3 of the "barrel" region representing the spared whisker. With over 10000 neurons per mouse, and up to 1000 active neurons in a single contiguous recording, these data provide a unique snapshot of a well-defined functional bottleneck necessary to perform a task. To answer our questions, we will first use unsupervised, objective algorithms to determine the size and number of cortical populations distributed across the L2/3 region. We will then use combined dimension reduction and decoding model approaches to work out what sensory or behavioural property is encoded by the joint activity of each population. Finally, we will track how identified populations emerge over task learning, along with their encoding properties.

Our results will the first to decompose a complete set of cortical populations, the first to show how multiple populations work together on a single computational task, and the first to show how multiple populations are recruited by learning, in what order, and how this predicts behaviour. Together, these will place strong constraints on theories of cortical coding and computation, and the interpretation of cortical activity data.

Clinical impact:
The dysfunction of cortical populations has implicated in both neurodegenerative (e.g. Alzheimer's disease) and mental health disorders (e.g. schizophrenia). The potentially wide ranging impact for understanding how cortex codes and computes with populations of neurons has driven an extraordinary research agenda. Even the highly-specific terms "cortical coding/computation" on PubMed yields more than 400 papers per year since 2013. Yet to date the vast majority of experimental studies have only been able to address a single such population. Our proposed work on deconstructing a complete set of cortical populations necessary for a specified task thus has the long term potential to advance theory and practice in a wide range of clinical settings. One broad arena of mid-term impact is likely to be on areas of cognitive neuroscience and neurology that use non-invasive imaging (fMRI, EEG and MEG) to study cortical function in humans.

Beyond neurological disorders, a potential short-term clinical impact is in the design of Brain Machine Interfaces. In human patients and primate test subjects, the cortical control of robot arms and prosthetic devices is achieved from decoders of recorded activity of small populations. These decoders assume a single population. Our results will likely inform the design of better decoders, and thus advance the ability to control any brain-machine interface that is dependent on cortical activity.


Economic & societal impact:
Recent success in neuro-inspired computation by private companies have captured the public imagination. The UK-based DeepMind company recently demonstrated a neural architecture that could learn by itself to outperform skilled human players on a range of video games, and to beat expert human players at Go. Such leaps in machine-learning have been built on foundational experimental work in neuroscience on the coding and computation by neural populations. Our work promises to open up new ideas of how cortical populations work, with the potential to inform new advances in cutting-edge, commercial machine learning.

The design of "neuromorphic" chips has become a primary goal of major computer processor companies (Qualcomm, IBM), as the leading candidate for the next-generation processor. The Manchester-based SpiNNaker project is at the forefront of this revolution. These general purpose processors aim to harness the immense computing power of brains by explicitly computing using populations of single neurons. As our proposed work will show for the first time how multiple cortical populations compute, it will strongly inform the design of computing paradigms for these chips.

Researcher career development:
Our proposed work programme contains a number of specific elements for the career development of the named 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 / cluster computing, development of machine learning technique, and application of advanced statistics. The named researcher will also gain experience in project management and supervision of students, enhancing future fellowship and tenured position applications.