近年来,越来越多的人工智能方法在解决传统自然科学等问题上大放异彩, 在一些重要的学科问题(例如蛋白质结构预测)上取得了令人瞩目的进展。在物理领域的研究中,非常多的物理问题都会涉及建模物体的的一些几何特征,例如空间位置,速度,加速度等。这种特征往往可以使用几何图这一形式来表示。不同于一般的图数据,几何图一个非常重要的特征是额外包含旋转,平移,翻转对称性。这些对称性往往反应了某些物理问题的本质。因此,最近以来,大量工作利用了几何图中的对称性,基于经典图神网络设计了很多具有等变性质的模型去解决对几何图建模问题。尽管在这一领域,等变图神经网络模型取得了长足的发展,但是还缺乏一个系统性的对这一领域的调研。为此,腾讯 AI Lab, 清华 AIR & 计算机系在综述:《Geometrically Equivariant Graph Neural Networks: A Survey》中,对等变图神经网络的结构和相关任务进行了一个系统梳理。
![Image](https://image.jiqizhixin.com/uploads/editor/4cc168d7-facc-4d31-98a3-746c9b45e8d7/640.png)
![Image](https://image.jiqizhixin.com/uploads/editor/f4646ecf-029e-4289-802c-c9b88f5099cd/640.png)
![Image](https://image.jiqizhixin.com/uploads/editor/915dd9a7-363d-4ddd-b654-3bba283704b9/640.png)
![Image](https://image.jiqizhixin.com/uploads/editor/9c976357-6e5e-42d6-b4ed-4bac421d8fc9/640.png)
![Image](https://image.jiqizhixin.com/uploads/editor/504bd723-39a4-43b9-adda-f5b4cfbdf5f1/640.png)
![Image](https://image.jiqizhixin.com/uploads/editor/dcd86e6d-f826-4349-9171-f5f3ad38280d/640.png)
![Image](https://image.jiqizhixin.com/uploads/editor/a20a4424-5095-48c5-a82f-6791119f6f14/640.png)
![Image](https://image.jiqizhixin.com/uploads/editor/eeab04a5-3fbc-4947-ac34-8d8ac29e097f/640.png)
![Image](https://image.jiqizhixin.com/uploads/editor/0991a634-d0f9-4669-98e8-2fb6379faee6/640.png)
![Image](https://image.jiqizhixin.com/uploads/editor/6d25a661-7737-4e94-9109-4c06517a1ecb/640.png)
![Image](https://image.jiqizhixin.com/uploads/editor/e4f245db-253b-43cd-a60c-b4846a0e8fc6/640.png)
![Image](https://image.jiqizhixin.com/uploads/editor/86971136-033b-453c-b966-4310a3c9ef57/640.png)
![Image](https://image.jiqizhixin.com/uploads/editor/49b9bdc2-72d8-4034-aacf-744d53f06e54/640.png)
![Image](https://image.jiqizhixin.com/uploads/editor/c669ce8b-ff48-44a9-8b8c-8e5cf3e15c66/640.png)
![Image](https://image.jiqizhixin.com/uploads/editor/bd0357ea-e29e-448e-a39c-457fa09a0fd2/640.png)
![Image](https://image.jiqizhixin.com/uploads/editor/b3ef8ecf-131c-4cbd-bace-b1e8ee75b890/640.png)