Experimental neuroscientists can record the signals communicated among the neurons that are collectively involved in producing meaningful behaviors, but making sense of these patterns of activity in terms of specific mental functions is challenging. This research project aims to discover the unseen mental processes that underlie such meaningful behavior from those recordings. The technology developed in this endeavor will uncover new ways of understanding mental processes hidden deep in the noisy signals collected from multiple neurons and will be used to derive new theoretical models (cognitive algorithms) to explain how populations of neurons work together. Such models will contribute to the development of diagnostic tools and neural prosthetics for cognitive dysfunctions in perception, working memory, and decision making, and can also inspire advances in machine learning and artificial intelligence.
The technical goal of this project is to develop a data-driven framework amenable to visualization and interpretation of neural activity underlying cognition. The core of the project is the identification and recovery of an interpretable low-dimensional nonlinear continuous dynamical system that underlies observed neural time series, and its validation through experimental perturbations. This will answer two key scientific questions: (1) How are task and cognitive variables represented in low-dimensional neural trajectories; and (2) What are the laws that govern the time evolution of the neural states. Answering these questions will help us understand how subjects implement and switch between different cognitive strategies, and more importantly, will provide a means for testing previously proposed theoretical models of the neural computations underlying cognition. This project will develop a number of statistical methods that can (i) extract private and shared noise from single-trial electrophysiological observations, (ii) combine recordings from multiple sessions to infer a common cognitive neural dynamics model, and (iii) design control stimulation to perturb the current neural state. Specifically, these tools will be applied to recordings from cortical areas involved in visuomotor decision-making to discover (1) how the co-variability in a population of sensory neurons encodes decision variables, (2) how the cognitive strategy changes when sensory evidence statistics change, and (3) the underlying dynamics that sustain spatial working memory. The success of this project could transform how the field analyzes population activity with low-dimensional structure in the context of cognitive tasks and beyond.
