Reproduce “Inferring networks from time series: a neural approach” on OpenResearch

About this paper

Thomas Gaskin, Grigorios A. Pavliotis, Mark Girolami

Researchers from Imperial College London and the University of Cambridge developed a neural network-based method for inferring large network adjacency matrices from time series data, providing principled uncertainty quantification. This approach scales to networks with 1000 nodes, inferring millions of potential edge weights in minutes on a GPU or hours on a CPU, and demonstrates up to an order of magnitude higher accuracy than traditional methods in noisy and underdetermined conditions.

Key insights

• The neural parameter calibration approach is generalizable beyond network inference, applicable to a broad class of high-dimensional parameter estimation problems.
• Uncertainty quantification can effectively reflect both observation noise and the degree to which an inference problem is underdetermined, by evaluating the likelihood of various network structures during training.
• Integrating a neural network with a differentiable forward model allows for efficient optimization through backpropagation, enabling scalability for large network inference problems.

Summary by alphaXiv.