Preparing Lessons for Progressive Training on Language Models

TL;DR

The paper introduces Apollo, a novel training method that enables efficient progressive expansion of language models by learning high-layer functions during low-layer training, significantly reducing resource use and environmental impact.

Contribution

Apollo is a new approach that prepares lessons for model expansion using low-value-prioritized sampling, weight sharing, and interpolation, improving training efficiency without pretrained models.

Findings

Achieves state-of-the-art acceleration ratios.

Rivals pretrained model-based methods in efficiency.

Reduces training time and environmental costs.

Abstract

The rapid progress of Transformers in artificial intelligence has come at the cost of increased resource consumption and greenhouse gas emissions due to growing model sizes. Prior work suggests using pretrained small models to improve training efficiency, but this approach may not be suitable for new model structures. On the other hand, training from scratch can be slow, and progressively stacking layers often fails to achieve significant acceleration. To address these challenges, we propose a novel method called Apollo, which prep\textbf{a}res lessons for ex\textbf{p}anding \textbf{o}perations by \textbf{l}earning high-\textbf{l}ayer functi\textbf{o}nality during training of low layers. Our approach involves low-value-prioritized sampling (LVPS) to train different depths and weight sharing to facilitate efficient expansion. We also introduce an interpolation method for stable model depth extension. Experiments demonstrate that Apollo achieves state-of-the-art acceleration ratios, even rivaling methods using pretrained models, making it a universal and efficient solution for training deep models while reducing time, financial, and environmental costs.