Dr. Gerald Pao（沖縄科学技術大学院大学 生物の非線形力学データサイエンス研究ユニット 准教授）

2025年03月03日(月) 16:00 より 17:30 まで

基生研 明大寺地区1階 第1セミナー室 (132-134)

超階層生物学センターバイオイメージング解析室　上野直人（7570）

The advent of Big Data has made it possible to address important challenges using deep learning approaches. Many problems that defied solutions using traditional physics-based strategies are soluble with artificial neural networks that are universal function approximators. However, the lack of explainability becomes a significant impediment to scientific understanding. We will provide a framework that creates mathematical tools for explainable science of complex systems. Here we provide examples from systems biology and neuroscience to understand networks as dynamical systems. We feature here the understanding of gene expression networks as a first example. Here we also show that using nonlinear dimensionality estimation methods based on delay embeddings, that are not based on correlation for the analysis of whole brain activity in Drosophila, zebrafish, mice and humans one finds that the dimensionality of brain activity is much lower than thought and that activity can be described on the surfaces of low dimensional manifolds. This has now been taken to single cell resolution where we show that the vast majority of neurons have dynamics that have the properties of low dimensional manifolds. This approach allows us to create predictive models of behavior based on brain activity and ultimately create models of brain activity that map brain activity to behavior at single neuron resolution. These computational abstractions largely meet the criteria of what would be understood as downloading the brain from experimental observations in animals. When run forward in time we obtain simulations of whole brain activity as well as behavior that closely resemble the original organism.