ROR: Dynamically switch optimizers during training

RichmanRonald · x · 2026-08-20

The paper 'Many Optimizers But Only One Training Path' proposes 'Repeated Optimizer Resampling' (ROR). Unlike fixing one optimizer, ROR dynamically searches for the best one during training. Every b epochs, candidate optimizers scout for s epochs, and the best one finishes the segment. Experiments on MNIST, Fashion-MNIST, and insurance data show one-epoch ROR uses 24%-35% of the training cost of exhaustive search while matching the best fixed optimizer's performance.

Original post →

More from Research

Research channel →