The vanishing gradient problem for stiff neural differential equations.

Gradient-based optimization of neural differential equations and other parameterized dynamical systems fundamentally relies on the ability to differentiate numerical solutions with respect to model parameters. In many stiff systems, it has been obser…

Preview only. Read the full abstract at the source