Preparing Lessons for Progressive Training on Language Models

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 prepares lessons for expanding operations by learning high-layer functionality 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.

Authors

• Yu Pan Ximalaya Inc.
• Ye Yuan University of Electronic Science and Technology of China
• Yichun Yin Huawei Noah’s Ark Lab, Shenzhen, Guangdong, China
• Jiaxin Shi Cloud BU, Huawei Technologies
• Zenglin Xu Fudan University
• Ming Zhang
• Lifeng Shang Huawei Noah’s Ark Lab, Shenzhen, Guangdong, China
• Xin Jiang School of Electronic Engineering, Xidian University, Xi’an, China
• Qun Liu Huawei Noah's Ark Lab

Keywords

• NLP: (Large) Language Models
• NLP: Learning & Optimization for NLP