Icml2022 | scalable depth Gaussian Markov random field

Thesis link ：https://arxiv.org/abs/2206.05032

The machine learning method on graph has been proved to be useful in many applications , Because they can handle general structured data . Gauss Markov random field (GMRFs) The framework provides a principled approach , Using the sparsity structure of graph to define Gaussian model . This article is in depth GMRF Based on the multi-layer structure , This paper presents a flexible method for general graphs GMRF Model , The model was initially proposed only for lattice graphs . By designing a new layer , We enable the model to be scaled to a large graph . The construction of this layer allows the use of variational reasoning and graphical neural networks of existing software frameworks for effective training . For Gaussian likelihood , The potential field can be inferred by Bayes with near accuracy . This allows predictions to be made , It is accompanied by uncertainty estimation . Experiments on a large number of synthetic and real data sets verify the effectiveness of the proposed model , In these experiments , It is better than other Bayesian and deep learning methods .