Accelerating Continuous Normalizing Flow with Trajectory Polynomial Regularization

In this paper, we propose an approach to effectively accelerating the computation of continuous normalizing flow (CNF), which has been proven be a powerful tool for tasks such as variational inference and density estimation. The training time cost CNF can extremely high because required number function evaluations (NFE) solving corresponding ordinary differential equations (ODE) is very large. We think that NFE results from large truncation errors ODEs. To address problem, add regularization. regularization penalizes difference between trajectory ODE its fitted polynomial regression. will approximate function, thus error smaller. Furthermore, provide two proofs claim additional does not harm quality. Experimental show our proposed method result in 42.3% 71.3% reduction on task estimation, 19.3% 32.1% auto-encoder, while testing losses are affected.