STEER: Simple Temporal Regularization For Neural ODEs
This addresses a computational bottleneck for researchers and practitioners using Neural ODEs, offering an incremental improvement that is easy to implement and compatible with existing methods.
The paper tackles the high computational cost of training Neural ODEs by proposing a simple regularization technique that randomly samples the end time of the ODE during training, which significantly decreases training time and improves performance across tasks like normalizing flows, time series models, and image recognition.
Training Neural Ordinary Differential Equations (ODEs) is often computationally expensive. Indeed, computing the forward pass of such models involves solving an ODE which can become arbitrarily complex during training. Recent works have shown that regularizing the dynamics of the ODE can partially alleviate this. In this paper we propose a new regularization technique: randomly sampling the end time of the ODE during training. The proposed regularization is simple to implement, has negligible overhead and is effective across a wide variety of tasks. Further, the technique is orthogonal to several other methods proposed to regularize the dynamics of ODEs and as such can be used in conjunction with them. We show through experiments on normalizing flows, time series models and image recognition that the proposed regularization can significantly decrease training time and even improve performance over baseline models.