In our lab, we use the tools of theoretical and computational neuroscience to study how behaviors are learned and how this learning is implemented in neural circuits.

Learning new behaviors requires the shaping of time-varying patterns of neural activity. The scientific questions that we seek to address focus on how neurons in the cerebral cortex interact with subcortical brain areas to learn things such as selecting actions to perform or how to perform a new motor skill. Understanding how such learning occurs is challenging because multiple brain areas are involved and because such behaviors may involve multiple timescales, from low-level limb movements to the cognitive level of goal-driven planning. By developing models of the subcortical and cortical neural circuits that control behavior, using data from experimental collaborators to inform and test these models, and drawing on recent advances in deep learning and artificial intelligence, our research aims to advance understanding of how learned motor behaviors are implemented in the brain.

The core of our research is about building mechanistic models to understand computations performed by the brain. Neuroscience has a decades-long history of drawing on ideas and models from theoretical physics and applied mathematics. These models make use of tools from areas including statistical mechanics, random matrix theory, dynamical systems, spin glass theory, stochastic processes, optimal control theory, and reinforcement learning. An example of a question that we have recently studied using some of these techniques: Why do behaviors become more automatic the more we practice them, and which mechanisms in the brain make this possible?

Our research also makes extensive use of ideas from deep learning and artificial intelligence. Modern approaches to AI are based on training neural networks to perform tasks ranging from face recognition to playing video games. A tremendous amount of AI research has focused on characterizing the architectures and algorithms that enable neural networks to solve these sorts of tasks. Much of our research takes inspiration from these advances and applies them to questions about the architectures of neural circuits in the brain and the algorithms that they use to learn. An example question: If we model model motor cortex as a recurrently connected neural network, what are the rules for how to change the synaptic strengths between neurons in a way that will make the neural dynamics do something useful (like telling an arm to reach for a cup of coffee)?

Because neuroscience is at its core an experimental field, a major focus of our work is to use the theoretical tools to address experimental data. Experimental neuroscience has entered an exciting new phase in recent years, with new technologies making it possible to record from and manipulate increasingly large populations of neurons in awake behaving animals. In theoretical neuroscience, we work in a feedback loop with experimental collaborators, in which our role is to use modeling and data analysis to interpret data and to contribute to experimental design by generating predictions that can be tested in future experiments. An example question: By observing the activity of neurons in the brain as a new task is learned, can we use models to infer the algorithms by which the synaptic strengths between neurons are changed in order to effect learning?