Adaptive Decoding of Noisy, Non-stationary Neural Spiking Activity

University Of Connecticut, Storrs CT

Investigators

Abstract

Decoding techniques allow neuroscientists to measure to what extent properties of external stimuli or movements can be reconstructed from neural activity observed in a specific brain area. Although most decoders only take average neural responses into account, decoding performance can also be affected by the variability in neural responses. Neural variability, in addition to a neuron’s mean response, depends on and can provide information about the external world. Here the researchers propose developing new statistical techniques that can more accurately model variability in neural activity and adapt to changes in neural response properties over time. More accurate decoding techniques have the potential to improve brain machine interfaces, such as motor prostheses and cochlear implants. The statistical tools developed here may also have applications outside of neuroscience as general methods for describing count data. As part of this project, graduate and undergraduate students participating in the research will receive interdisciplinary training in computational neuroscience and gain familiarity with the underlying neural systems. The models and results from this project will be integrated into graduate computational neuroscience coursework, and all code will be shared publicly with the neuroscience community to reproduce and extend our results. The primary aim of this work is to develop model-based statistical approaches that can more accurately describe within-trial and between-trial variability in neural spiking activity. The researchers will apply these models to 1) characterize noise and non-stationarity in single neuron spiking, 2) evaluate how variability is shared in populations of neurons, and 3) more accurately and robustly decode external variables. This work is based on a framework that combines Conway-Maxwell-Poisson models of spike count noise and adaptive filtering models of non-stationarity. Since most existing methods ignore non-Poisson noise, non-stationarity, or both, the work proposed here will provide a new set of tools for systems neuroscience. The research will examine multi-electrode spike train recordings from multiple existing data sets and quantify heterogeneity in neural responses across cell types, behaviors, and brain areas. There is growing evidence that neural variability depends on many factors, including an animal’s alertness and motivation, as well as properties of the environment and task. The results of this modeling will provide a more detailed understanding of the neural code and the role of variability in this code. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.