Sjfu

Surgical Hill-Kinetics Saturation: Eliminating Finite-Time Blowup and Adaptive Solver Grid Collapse in Neural ODEs

Continuous-Time Neural Networks and Neural Ordinary Differential Equations (Neural ODEs) suffer from severe numerical stiffness, finite-time gradient blowup, and adaptive solver grid collapse (NFE -> infinity) when weight magnitudes accumulate during training. In this work, we pr...