Automated Model Inference from Neural Dynamics for a Mechanistic Understanding of Cognition

Descripción del proyecto

Un modelo basado en el aprendizaje automático para el estudio exhaustivo de la cognición

La introducción de modelos mecanicistas bien alineados con los datos experimentales podría mejorar considerablemente la comprensión de las propiedades celulares y de red que subyacen a la cognición, así como de lo que conduce a condiciones patológicas. Financiado por las Acciones Marie Skłodowska-Curie, el equipo del proyecto AutoMIND pretende desarrollar una herramienta de inferencia de modelos basada en el aprendizaje automático capaz de identificar los parámetros de modelos particulares de redes neuronales de impulsos, al tiempo que captura dinámicas neuronales objetivo arbitrarias a partir de registros del encéfalo humano. La herramienta aprovecha los recientes avances en las técnicas de inferencia basadas en la simulación e incorpora el aprendizaje múltiple de parámetros simultáneos y las simulaciones basadas en el gradiente. Además, es capaz de resolver problemas neurocientíficos mediante experimentos «in silico».

Objetivo

Human cognition depends on complex coordinated dynamics of neural populations, which is shaped by a rich heterogeneity in cellular properties and network connectivity patterns of neural circuits. While cognitive neuroscience leverages macroscopic brain signals to relate neural activity to behavioral states, we currently cannot dissect them for their physiological contributions, hindering mechanistic interpretations of experimental data. One way to systematically study how physiological parameters shape neural dynamics is through mechanistic modeling of spiking neural networks. However, current modeling approaches are not quantitatively constrained by observed electrophysiological data, and often require painstaking and ad-hoc parameter-tuning by hand. Efficient discovery of mechanistic models that are consistent with experimental data would dramatically accelerate our understanding of how cellular and network properties impact cognition, and why it breaks down in pathological states, representing a radical departure from how neural data is analyzed in cognitive neuroscience. To this end, I propose to develop a machine learning-assisted model inference tool—Automated Model Inference from Neural Dynamics (AutoMIND)—that can identify parameters of candidate spiking neural network models that could capture arbitrary target neural dynamics from human brain recordings. AutoMIND extends on recent advances in simulation-based inference techniques, incorporating simultaneous parameter manifold-learning and gradient-based simulations. AutoMIND has broad utility for tackling neuroscientific questions by enabling expedited in-silico experiments. Here, I apply it to multiscale neural data to study how cellular and network properties shape: 1) the emergence of synchronous network oscillations during early neurodevelopment, and 2) the difference in neural dynamics and computation between sensory and association cortices—two questions of fundamental importance to neuroscience.