Inferring structure in neural time series data: dynamics and connectivity

The ability to derive insights from complex high-dimensional data, such as neural data, is important to improve our understanding of the underlying system. In this thesis, we approach this by studying two aspects of neural data: dynamics and connectivity. First, we analyze the dynamics of the sponta...

Full description

Saved in:
Bibliographic Details
Main Author: Suryadi
Other Authors: Chew Lock Yue
Format: Thesis-Doctor of Philosophy
Language:English
Published: Nanyang Technological University 2023
Subjects:
Online Access:https://hdl.handle.net/10356/170411
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:The ability to derive insights from complex high-dimensional data, such as neural data, is important to improve our understanding of the underlying system. In this thesis, we approach this by studying two aspects of neural data: dynamics and connectivity. First, we analyze the dynamics of the spontaneous activity in the habenula, a region in the brain known to be involved in a wide range of functions. We found that the habenula is situated in the reverberating regime, a state near but not at criticality, which balances the benefits and drawbacks of criticality in information processing. Subsequently, we approach the study of connectivity through Granger causality. We develop Jacobian Granger causality (JGC), a neural network-based approach towards nonlinear Granger causality that works well with high-dimensional data, is stable with respect to regularization, and performs on par or even exceed that of the current state of the art across different settings: the inference of Granger causal variables, interaction time lag, and interaction sign. We subsequently extend JGC towards sparse count data to facilitate network inference on neural spike data. We show that it outperforms a recent competing approach, is reliable in discriminating between excitatory and inhibitory connections, and apply it to infer the network from real spike train data from monkey visual cortex when presented with two stimuli: white noise and natural movie. Among other findings, JGC discovers the existence of bursting neurons, and in particular a larger set of bursting neurons in the natural movie case compared to white noise. This corresponds with literature where bursting activity is known to encode salient visual information, which is present in natural scenes.