6-1 科学的图神经网络.pdf

1、单击此处编辑母版标题样式GNN for Science黄文炳清华大学智能产业研究院目录1.1.背景介绍背景介绍2.相关研究3.最新进展4.总结目录目录人工智能蓬勃发展2020-5OpenAI 发布GPT32020-6Google发布ViT2021-12021-6智源发布悟道2.02021-2OpenAI 发布图像版GPT3Google发布Switch Transformer2021-7DeepMind 连登Nature2021-8AI for Science：从人工智能到人工专家智能DeepMind利用神经网络提升DFT关于电子相互作用的预测，登上ScienceAdvancing mathem

2、atics by guiding human intuition with AI.2021-12Magnetic control of tokamak plasmas through deep reinforcement learning.2022-02Pushing the frontiers of density functionals by solving the fractional electron problem.2021-12DeepMind利用AI发现数学新见解加速数学证明，登上Nature封面DeepMind和EPFL合作利用强化学习算法控制核聚变，登上NatureAI+数学

3、AI+物理AI+物理化学图（Graph）广泛存在于科学领域分子结构蛋白质-药物结合多体运动抽象成图结构抽象成图结构图神经网络（GNN）是分析图结构的有效工具之一Sperduti,Alessandro and Starita,Antonina.1997Sperduti,Alessandro,and Antonina Starita.Supervised neural networks for the classification of structures.1997.意大利学者首次定义“GNN”Scarselli et al.2005年最先定义“GNN”Gori,Marco,Gabriele M

4、onfardini,and Franco Scarselli.A new model for learning in graph domains.Proceedings.2005 IEEE International Joint Conference on Neural Networks,2005.Vol.2.IEEE,2005.Scarselli,Franco,et al.Graph neural networks for ranking web pages.The 2005 IEEE/WIC/ACM International Conference on Web Intelligence(

5、WI05).IEEE,2005.Scarselli,Franco,et al.The graph neural network model.IEEE Transactions on Neural Networks 20.1(2008):61-80.Lecun团队最先在机器学习会议上引入图卷积概念提出了两种不同的图卷积网络：1.空间卷积；2.谱卷积ICLR，2014过去几年，GNN的研究得到了越来越多的关注Joan Bruna et al.NIPS 201405010015020025030035020142015201620172018201920202021Number of GNN Pap

6、ersICLRICMLNeurIPSKDDAllGNN1.0：Understanding GNN as RNNBefore 2000Sperduti,Alessandro,and Antonina Starita.(TNN 97)propose the generalized recursive neuron for the graph classification problem on Trees/DAGs.This generalized recursive neuron can only generate the graph representations.From 2000 to 20