Conference agendas and presentations from major AI conferences like NeurIPS, ICML, etc.

Apple and EPFL Introduce AdEMAMix, a Novel AI Optimizer with 1.95x Improvement

Sep 10, 2024, 01:11 AM

Researchers from Apple and EPFL have introduced AdEMAMix, a novel optimization approach leveraging dual exponential moving averages to enhance gradient efficiency and improve large-scale model training performance. The new optimizer, which operates with just 120 lines of code, claims a 1.95x improvement over the widely-used AdamW optimizer. AdEMAMix requires 95% fewer training tokens than AdamW to reach the same level of performance. The approach utilizes two exponential moving averages for the numerator of Adam, a fast one with a low beta and a slow one with a high beta, which could explain its superior performance in various optimization scenarios, including FedOpt variants like DiLoCo.