KO: Kinetics-inspired Neural Optimizer with PDE Simulation Approaches.

arXiv:2505.14777v2 Announce Type: replace-cross Abstract: The design of effective optimization algorithms for neural networks remains a fundamental challenge, and most existing methods rely on heuristic extensions of gradient-based updates. We introduce KO (Ki