Paper ID: 2410.05021

DEPT: Decoupled Embeddings for Pre-training Language Models

Alex Iacob, Lorenzo Sani, Meghdad Kurmanji, William F. Shen, Xinchi Qiu, Dongqi Cai, Yan Gao, Nicholas D. Lane

Language model pre-training benefits from diverse data to enhance performance across domains and languages. However, training on such heterogeneous corpora requires extensive and costly efforts. Since these data sources vary lexically, syntactically, and semantically, they cause negative interference or the ``curse of multilinguality''. We propose a novel pre-training framework to alleviate this curse. Our method, DEPT, decouples embeddings from the transformer body while simultaneously training the latter in multiple contexts. DEPT enables training without a shared global vocabulary and: (1) can train robustly and effectively under significant data heterogeneity, (2) reduces token embedding parameters by up to 80% and the communication costs by 675x for billion-scale models, (3) enhances model generalization and plasticity in adapting to new languages and domains, and (4) permits training with custom optimized vocabularies per data source. We demonstrate DEPT's potential via the first vocabulary-agnostic federated multilingual pre-training of a 1.3 billion-parameter model, limiting its embedding size to 102.4 million instead of 512 million.

Submitted: Oct 7, 2024