guest ::
| Home > Publications database > Analysis of Electrophysiological Data with Invertible Neural Networks |
| Home > Publications database > Analysis of Electrophysiological Data with Invertible Neural Networks |
| Book/Dissertation / PhD Thesis | FZJ-2026-03629 |
2026
Forschungszentrum Jülich GmbH Zentralbibliothek, Verlag
Jülich
ISBN: 978-3-95806-945-9
This record in other databases:
Please use a persistent id in citations: doi:10.34734/FZJ-2026-03629
Abstract: Understanding how neural populations represent information remains a central challenge in neuroscience. With highly parallel neural recordings from many neurons, a key question arises: how can we extract meaningful structure from complex population dynamics? This thesis develops new approaches for analyzing neural data, focusing on temporal precision of spike-timing correlations and higher-order correlation structure in continuous population activity. Groups of neurons exhibit highly structured, precisely timed correlations despite individual responses appearing variable. The fundamental question becomes: are observed correlations merely byproducts of firing rate modulations, or do they reflect genuine temporal coordination that provides additional computational power? Population-level neural activity typically resides on low-dimensional manifolds within high-dimensional state spaces. Traditional methods like Principal Component Analysis are constrained by linearity and Gaussianity assumptions, unable to reveal curved manifolds and nonlinear dependencies crucial for understanding neural computation. This thesis tackles both challenges through four interconnected studies spanning discrete to continuous neural signals. The first chapter advances analyzing precisely timed correlations through exhaustive assessment of surrogate generation approaches, offering recommendations for method selection tailored to specific experimental objectives. The second chapter examines how recurrent circuit dynamics transform neural representations, introducing a mean-field theory that distills complex network behavior into three critical variables: population-averaged activity and correlation measures quantifying response variability within and between stimulus categories. The third chapter develops methods for extracting hidden interdependencies using invertible neural networks, demonstrating techniques for identifying microscopic interactions directly from observed data without strong prior assumptions. The fourth chapter presents a comprehensive framework for neural manifold analysis using invertible networks. We implement novel extensions including a specialized loss function, a multimodal latent space, and an analytical representations enabling direct computation of higher-order statistics and curvature measures. Analysis of macaque visual cortex recordings reveals that distinct behavioral contexts produce manifolds with characteristic geometric properties. These innovations provide new capabilities for deciphering how neural systems encode and process information, advancing our understanding of computational principles governing brain function.
; 
|
The record appears in these collections: |
PDF