Xaq Pitkow - US grants

A primary goal of neuroscience is to understand how the brain works-- not in artificial lab tasks, but when using its full capabilities to thrive in the rigors of the natural environment. Neuroscience has made enormous progress by examining how the brain performs simplified tasks, but these tasks do not expose the richly adaptive dynamics that the brain must use in a changing world. Therefore, the current neuroscientific understanding of the brain is missing fundamental ingredients. The current project begins to fill this gap, providing a new paradigm for the conduct of behavioral neuroscience and offering an unprecedented opportunity to observe the neural computations that solve a complex natural task. Team members will record activity of many neurons in multiple areas of a mouse brain while the mouse is foraging in a virtual reality environment, and develop mathematical models to make sense of the complex data. This research will thereby provide a unique training opportunity for undergraduate and graduate students in both computational and experimental neuroscience. The project results will be widely disseminated by sharing data, computational models, and analysis techniques with the neuroscience community through public data repositories, so conclusions can be replicated and extended. This research will thereby advance society?s goals of understanding the biology of healthy and disordered brains, with the ultimate hope of repairing neurological problems.

Experimenters will train mice to forage in a virtual reality environment, while recording activity from many neurons in four brain areas involved in vision and navigation: visual cortex, entorhinal cortex, posterior parietal cortex and hippocampus. State-of-the-art analysis techniques will be used to describe the mouse's behavior, and to discover neural representations of the internal models that express the animal's beliefs about things that cannot be observed directly in sense data. Finally, the project will uncover how neural representations of critical task variables are communicated and transformed across brain areas, guided by the hidden variable dynamics of the behavioral model. Together, these experiments, theory, and analysis will provide an unprecedented, system-wide understanding of neural computation, ranging from the scale of individual neurons up to a multi-region system. A key quality of the approach is the pervasive influence of theory, both in structuring experiments and dictating analyses. Since the great strength of the human brain is its ability to comprehend the hidden structure in the world, this approach takes an essential step toward unraveling the mysteries of cognition.

This project aims to understand how large populations of neurons transform their encoded information to drive behaviors meaningful to the organism. This will be accomplished in two ways. First, the research will derive new analysis methods that experimentalists can use to interpret neural data from naturalistic tasks of moderate complexity. Second, by the project will create a broadly applicable computational framework for synthesizing these analyses into a theory of probabilistic neural computation. Both of these components are informed by three basic principles: information in the brain is distributed across many neurons, sensory evidence is weighted by its reliability, and neural computation occurs in multiple stages. Current analyses that connect animal behavior to neural activities apply to tasks that are so simple that an animal would not actually need a brain to solve them: the same computations could be accomplished in a single step by wiring the sensory organs directly to the muscles. Clearly there is a need to study more complex tasks that require multi-step computations, and the proposed research will provide the rigorous statistical foundation needed to analyze data from such studies. The research will also have a broader educational impact by creating interactive teaching games that explain concepts needed for thinking about big neuroscience data.

The long-term goal of this research program is to explain brain function by constructing quantitative theories of how distributed nonlinear neural computation implements principles of statistical reasoning. To accomplish this goal, this project will create a normative theory for what information about naturalistic tasks should be encoded in neural populations, and data analyses that can reveal which aspects of that information are actually decoded. The normative theory is based on probabilistic population codes, a model in which large-scale neural activity patterns encode not just estimates of a stimulus, but also the reliability of those estimates. This model is currently applied only to small-scale inference problems, and one aim of this project is to extend this model by constructing biologically plausible network models for complex naturalistic tasks involving many interacting variables. The key components of this model, and indeed of any model of naturalistic computation, are nonlinear operations. To determine whether the posited nonlinear computations occur in a real brain, the other aim of the project is to derive a statistical analysis technique centered on a novel generalization of standard choice-related activity, termed nonlinear choice correlation. By combining this measure with estimates of neural correlations, experimentalists will be able to infer the class of distributed nonlinear computations the brain uses from simultaneous recording of neural activity and animal behavior.

Scientists have learned much about how the activity of individual neurons relates to sensations, thoughts, feelings, and behavior. However, all neurons live and function within the enormous web that is the brain. Scientists are still far from understanding how the complex, dynamic dance between the cells that form this web makes us who we are. While new technologies allow scientists to observe the activity of massive numbers of neurons in living brains, they still cannot make sense of such data. The goal of this NeuroNex Theory team is to develop a new lens - a set of mathematical and computational tools - for making sense of the dynamic brain activity data. These techniques will allow them to deduce when and how observed neurons interact, how these interactions are altered by stimuli, and how they may govern behavior. The resulting combination of revolutionary experimental and mathematical tools will provide neuroscientists with unprecedented insights into the brain's distributed neural computations.

This NeuroNex Theory team will develop statistical tools that will provide insights into the computations performed by neuronal ensembles, by relating what the animal observes to recorded neural activity as well as how the neural activity relates to behavior. They will validate the methods using simulated neuronal networks, and apply them to recordings from the mouse brain. The team will also develop and test their approach using modern deep neural networks (convolutional nets and recurrent architectures) that achieve or exceed human performance in many hard tasks. This will allow them to quantify how interactions in these artificial networks change with the stimulus, and compare and contrast the results with data from animals. This is a broad set of goals that requires a combination of approaches and diverse expertise. The team therefore consists of experimental and theoretical neuroscientists, mathematicians, and statisticians who will work together closely. The true test of this work will be the impact it will have on brain science in general. The team is therefore committed to sharing their techniques, code and data. They have also assembled a group of End Users to be early adopters and testers of these methods. This NeuroNex Theory Team award is funded as part of the BRAIN Initiative and NSF's Understanding the Brain activities.