Residual network with self-adaptive time step size

Xiyuan Li, Xin Zou, Weiwei Liu

Abstract: Highlights•We bridge the relations between ResNet and ODE.•The convergence and stability of the residual function are analyzed by ODE.•We claim that the time step of ResNet depends on the feature map and parameters.•We propose AdaTS-ResNet made of feature map and parameters with small Lipschitz.•AdaTS-ResNet outperforms other methods designed based on the adaptive ODE in experiments.