EXTRACTING NEURONAL DYNAMICS AT HIGH SPATIOTEMPORAL RESOLUTIONS: THEORY, ALGORITHMS, AND APPLICATION

Abstract

Analyses of neuronal activity have revealed that various types of neurons, both at the single-unit and population level, undergo rapid dynamic changes in their response characteristics and their connectivity patterns in order to adapt to variations in the behavioral context or stimulus condition. In addition, these dynamics often admit parsimonious representations. Despite growing advances in neural modeling and data acquisition technology, a unified signal processing framework capable of capturing the adaptivity, sparsity and statistical characteristics of neural dynamics is lacking. The objective of this dissertation is to develop such a signal processing methodology in order to gain a deeper insight into the dynamics of neuronal ensembles underlying behavior, and consequently a better understanding of how brain functions. The first part of this dissertation concerns the dynamics of stimulus-driven neuronal activity at the single-unit level. We develop a sparse adaptive filtering framework for the identification of neuronal response characteristics from spiking activity. We present a rigorous theoretical analysis of our proposed sparse adaptive filtering algorithms and characterize their performance guarantees. Application of our algorithms to experimental data provides new insights into the dynamics of attention-driven neuronal receptive field plasticity, with a substantial increase in temporal resolution. In the second part, we focus on the network-level properties of neuronal dynamics, with the goal of identifying the causal interactions within neuronal ensembles that underlie behavior. Building up on the results of the first part, we introduce a new measure of causality, namely the Adaptive Granger Causality (AGC), which allows capturing the sparsity and dynamics of the causal influences in a neuronal network in a statistically robust and computationally efficient fashion. We develop a precise statistical inference framework for the estimation of AGC from simultaneous recordings of the activity of neurons in an ensemble. Finally, in the third part we demonstrate the utility of our proposed methodologies through application to synthetic and real data. We first validate our theoretical results using comprehensive simulations, and assess the performance of the proposed methods in terms of estimation accuracy and tracking capability. These results confirm that our algorithms provide significant gains in comparison to existing techniques. Furthermore, we apply our methodology to various experimentally recorded data from electrophysiology and optical imaging: 1) Application of our methods to simultaneous spike recordings from the ferret auditory and prefrontal cortical areas reveals the dynamics of top-down and bottom-up functional interactions underlying attentive behavior at unprecedented spatiotemporal resolutions; 2) Our analyses of two-photon imaging data from the mouse auditory cortex shed light on the sparse dynamics of functional networks under both spontaneous activity and auditory tone detection tasks; and 3) Application of our methods to whole-brain light-sheet imaging data from larval zebrafish reveals unique insights into the organization of functional networks involved in visuo-motor processing.