Stochastic Dynamical Systems Analysis of Neural Information Representations

Massachusetts General Hospital, Boston MA

Investigators

Abstract

PROJECT SUMMARY Stochastic Dynamical Systems Analysis of Neural Information Representations The nervous system is composed of specialized cells called neurons that transmit information through the pattern of electrical impulses known as action potentials or spikes. The sequence of spikes forms point process time-series, i.e. a series of ones and zeros. Neural systems use the firing patterns of spike trains to encode representations of relevant biological signals and external stimuli. An experimental system widely used to study neural information encoding is the rat hippocampus. The hippocampus is a brain region critical for the formation and storage of both short- and long-term memories. Within the hippocampus the rat uses neurons known as place cells to develop a spatial map of an environment. As an animal moves through its environment, a hippocampal place cell demarcates, within as few as 5 minutes, its own region in the environment by firing spikes only when the animal is within that region. The region of the environment in which the cell fires is termed its place field. Large numbers of hippocampal place cells tile each environment with overlapping place fields and their ensemble-firing pattern gives a continuous representation (memory) of the animal's location in space. How ensembles of place cell neurons work together to represent spatial information is an important question. An active area of neuroscience research is the development of mathematical techniques to decipher this ensemble encoding. Because neural spike trains are point processes, standard statistical signal processing techniques for continuous data have limited application in the analysis of neural systems. Accurate processing of neural system dynamics requires development of techniques to characterize correctly the point process nature of neural encoding. Martingale theory offers an efficient way to relate the current behavior of a point process stochastic dynamical system to the system's history and provides a prescription for finding the best estimate of signal encoded in a multidimensional point process time-series. While martingale theory has been applied in cancer survival studies, queuing theory and communication problems, it has yet to be adapted to signal processing in neural systems. Another signal processing discipline whose ideas have had limited application in neuroscience is adaptive estimation. Adaptive estimation offers a way to track the dynamics of how a neural system updates its encoding of a relevant biological stimulus. This project will study two problems in neural signal processing: 1) development adaptive estimation algorithms to track instantaneously the dynamics of spatial information encoding by individual rat hippocampal place cell neurons; and 2) development of neural spike train decoding algorithms based on martingale theory to study information representation by ensembles of rat hippocampal place cell neurons. This work will provide a statistical framework for analysis of experimental studies on the role the hippocampus plays in memory formation. The adaptive estimation and martingale methods developed in this research will offer a new statistical paradigm for studying the dynamics of information representation and transmission in neural systems.