Locally Regularized Neural Differential Equations: Some Black Boxes Were Meant to Remain Closed!
This work addresses computational efficiency for researchers and practitioners using implicit layer models, offering a flexible drop-in method without strict requirements, though it is incremental as it builds on existing regularization techniques.
The paper tackles the high computational cost of training Neural Differential Equations by using internal solver heuristics to guide learning toward easier-to-integrate dynamical systems, achieving reductions in function evaluations to 0.556-0.733x and prediction speedups of 1.3-2x.
Implicit layer deep learning techniques, like Neural Differential Equations, have become an important modeling framework due to their ability to adapt to new problems automatically. Training a neural differential equation is effectively a search over a space of plausible dynamical systems. However, controlling the computational cost for these models is difficult since it relies on the number of steps the adaptive solver takes. Most prior works have used higher-order methods to reduce prediction timings while greatly increasing training time or reducing both training and prediction timings by relying on specific training algorithms, which are harder to use as a drop-in replacement due to strict requirements on automatic differentiation. In this manuscript, we use internal cost heuristics of adaptive differential equation solvers at stochastic time points to guide the training toward learning a dynamical system that is easier to integrate. We "close the black-box" and allow the use of our method with any adjoint technique for gradient calculations of the differential equation solution. We perform experimental studies to compare our method to global regularization to show that we attain similar performance numbers without compromising the flexibility of implementation on ordinary differential equations (ODEs) and stochastic differential equations (SDEs). We develop two sampling strategies to trade off between performance and training time. Our method reduces the number of function evaluations to 0.556-0.733x and accelerates predictions by 1.3-2x.