DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter

TLDR

Transfer learning from large pre‑trained models is common in NLP, but running these models on edge devices or under limited computational budgets is challenging. The authors propose pre‑training a smaller general‑purpose language model, DistilBERT, that can be fine‑tuned with performance comparable to larger models. DistilBERT is pre‑trained using a triple loss that combines language modeling, distillation, and cosine‑distance objectives. DistilBERT achieves a 40 % reduction in size, retains 97 % of BERT’s language understanding, runs 60 % faster, is cheaper to pre‑train, and performs well in on‑device experiments.

Abstract

As Transfer Learning from large-scale pre-trained models becomes more prevalent in Natural Language Processing (NLP), operating these large models in on-the-edge and/or under constrained computational training or inference budgets remains challenging. In this work, we propose a method to pre-train a smaller general-purpose language representation model, called DistilBERT, which can then be fine-tuned with good performances on a wide range of tasks like its larger counterparts. While most prior work investigated the use of distillation for building task-specific models, we leverage knowledge distillation during the pre-training phase and show that it is possible to reduce the size of a BERT model by 40%, while retaining 97% of its language understanding capabilities and being 60% faster. To leverage the inductive biases learned by larger models during pre-training, we introduce a triple loss combining language modeling, distillation and cosine-distance losses. Our smaller, faster and lighter model is cheaper to pre-train and we demonstrate its capabilities for on-device computations in a proof-of-concept experiment and a comparative on-device study.