When and why PINNs fail to train from the viewpoint of neural tangent kernel
组织者
演讲者
时间
2023年11月17日 22:00 至 22:30
地点
Online
摘要
In this talk, we will share several literatures about the explanation of training dynamics of Physics-informed neural network. Especially, as a uniform framework, neural tangent kernel demonstrates the dynamical behavior when width of NN goes to infinity in the single layer setting. NTK illustrates training one-layer NN with infinite width is the same as a kernel regression. As a consequence, NTK will exhibit the training weights of different parts in loss function which will provide the guidance for adaptive algorithm of revised PINN.
演讲者介绍
焦小沛,本科毕业于上海交通大学致远学院,博士毕业于清华大学数学科学系。先后在北京雁栖湖应用数学研究院,荷兰特文特大学从事博士后工作。现研究方向包括有限维滤波理论,丘-丘滤波方法,物理信息神经网络以及生物信息学。研究兴趣主要集中于(1)利用李代数等几何工具进行偏微分方程求解与有限维滤波系统的分类;(2)设计基于物理信息神经网络的新型数值算法。