When Does Scale-Invariant Optimization Become Unstable? An Exact Schedule Law with Weight Decay
Researchers derive an exact law linking learning-rate schedules and weight decay in normalized networks, pinpointing when scale-invariant optimization destabilizes.