P-BERT: Hardware-Aware Optimization of BERT Using Evolutionary Techniques

Lim Sze Ying, Zhehui Wang, Rick Siow Mong Goh, Tao Luo

Keywords: Model Compression, Large Language Models, Computation Complexity, BERT, Hardware-Aware

TL;DR: P-BERT combines pruning, quantization, and knowledge distillation to cut BERT's computational needs by 60%, maintaining accuracy and scores.

Abstract: Transformer-based models have emerged as the go-to standards in Natural Language Processing (NLP), revolutionizing the landscape of NLP applications. As complex models continue to proliferate, the need for more efficient computational processing becomes increasingly imperative. This has led to the rise of model compression techniques, implemented to target computational inefficiencies. Expounding on this, we propose Pyramid-BERT (P-BERT), the integration of three established model compression techniques to further reduce the computational inefficiency of the standard BERT models, and subsequently optimize BERT under the hardware characteristics. Specifically, the techniques employed are pruning, quantization, and knowledge distillation. The first two aforementioned correlated techniques work simultaneously to remove redundant specifications while leveraging knowledge transfer from baseline models. These techniques enable a substantial reduction in computational cost, making P-BERT highly suitable for portable, low-power devices such as cellphones, wearable devices, and smartwatches, and thus enabling hardware-friendly processing on various computing engines. Additionally, we will be proposing a new metric, the inverted computational complexity to quantify the complexity and efficacy of the model. This metric aims to more accurately capture the hardware-specific performance characteristics. Our experimental results show that P-BERT achieves a remarkable reduction of at least 60\% in the inverted computational complexity ratio while ensuring comparable accuracy and scores across many downstream tasks compared with the baseline BERT models.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5609