LaProp authors resurface 2020 optimizer: momentum after normalization beats Adam on stability

Gradientdinner · x · 2026-08-25

Liu Ziyin and colleagues resurface LaProp (arXiv:2002.04839), their 2020 optimizer questioning whether momentum accumulation in Adam is "wrong". LaProp performs momentum accumulation after normalization, making training far more stable across parameter ranges.

The paper identifies a previously unrecognized problem of Adam-style optimizers: unnecessary coupling between momentum and adaptivity causes instability and divergence when the two parameter groups are mismatched. Decoupling grants greater hyperparameter flexibility and allows direct interpolation between signed-gradient and adaptive-gradient methods. Experiments show consistently improved speed and stability over Adam across tasks; the authors suggest it may still be useful today.

Original post →

More from Research

Research channel →