Headless Language Models: Learning without Predicting with Contrastive Weight Tying

TL;DR

This paper introduces Contrastive Weight Tying, a novel pretraining method for headless language models that reconstructs embeddings contrastively, reducing computational costs and improving downstream task performance.

Contribution

It presents a new contrastive pretraining approach for headless language models that significantly lowers training costs and boosts accuracy across multiple benchmarks.

Findings

Reduced training computational requirements by up to 20 times.

Achieved +1.6 GLUE score improvement.

Attained +2.7 LAMBADA accuracy increase.

Abstract

Self-supervised pre-training of language models usually consists in predicting probability distributions over extensive token vocabularies. In this study, we propose an innovative method that shifts away from probability prediction and instead focuses on reconstructing input embeddings in a contrastive fashion via Constrastive Weight Tying (CWT). We apply this approach to pretrain Headless Language Models in both monolingual and multilingual contexts. Our method offers practical advantages, substantially reducing training computational requirements by up to 20 times, while simultaneously enhancing downstream performance and data efficiency. We observe a significant +1.6 GLUE score increase and a notable +2.7 LAMBADA accuracy improvement compared to classical LMs within similar compute budgets.