Neuro-GSTH: A Geometric Scattering and Persistent Homology Framework for Uncovering Spatiotemporal Signatures in Neural Activity DOI Creative Commons
Dhananjay Bhaskar, Yanlei Zhang, Jessica L. Moore

и другие.

bioRxiv (Cold Spring Harbor Laboratory), Год журнала: 2023, Номер unknown

Опубликована: Март 24, 2023

Abstract Neural signals are high-dimensional, noisy, and dynamic, making it challenging to extract interpretable features linked behavior or disease. We introduce Neurospectrum , a framework that encodes neural activity as latent trajectories shaped by spatial temporal structure. At each timepoint, represented on graph capturing relationships, with learnable attention mechanism highlighting important regions. These embedded using wavelets passed through manifold-regularized autoencoder preserves geometry. The resulting trajectory is summarized principled set of descriptors - including curvature, path signatures, persistent homology, recurrent networks -that capture multiscale geometric, topological, dynamical features. drive downstream prediction in modular, interpretable, end-to-end trainable framework. evaluate simulated experimental datasets. It tracks phase synchronization Kuramoto simulations, reconstructs visual stimuli from calcium imaging, identifies biomarkers obsessive-compulsive disorder fMRI. Across tasks, uncovers meaningful dynamics outperforms traditional analysis methods.

Язык: Английский

Neuro-GSTH: A Geometric Scattering and Persistent Homology Framework for Uncovering Spatiotemporal Signatures in Neural Activity DOI Creative Commons
Dhananjay Bhaskar, Yanlei Zhang, Jessica L. Moore

и другие.

bioRxiv (Cold Spring Harbor Laboratory), Год журнала: 2023, Номер unknown

Опубликована: Март 24, 2023

Abstract Neural signals are high-dimensional, noisy, and dynamic, making it challenging to extract interpretable features linked behavior or disease. We introduce Neurospectrum , a framework that encodes neural activity as latent trajectories shaped by spatial temporal structure. At each timepoint, represented on graph capturing relationships, with learnable attention mechanism highlighting important regions. These embedded using wavelets passed through manifold-regularized autoencoder preserves geometry. The resulting trajectory is summarized principled set of descriptors - including curvature, path signatures, persistent homology, recurrent networks -that capture multiscale geometric, topological, dynamical features. drive downstream prediction in modular, interpretable, end-to-end trainable framework. evaluate simulated experimental datasets. It tracks phase synchronization Kuramoto simulations, reconstructs visual stimuli from calcium imaging, identifies biomarkers obsessive-compulsive disorder fMRI. Across tasks, uncovers meaningful dynamics outperforms traditional analysis methods.

Язык: Английский

Процитировано

3