[Paper Review] PoWER-BERT: Accelerating BERT Inference via Progressive Word-vector Elimination

Saurabh Goyal, Anamitra R. Choudhury+4 more|arXiv (Cornell University)|Jan 24, 2020

Topic Modeling52 citations

TL;DR

PoWER-BERT reduces BERT inference time by progressively eliminating redundant word-vectors (intermediate encoder outputs) guided by self-attention, achieving up to 4.5x speedup with <1% accuracy loss on GLUE tasks, and up to 6.8x on ALBERT with similar accuracy loss.

ABSTRACT

We develop a novel method, called PoWER-BERT, for improving the inference time of the popular BERT model, while maintaining the accuracy. It works by: a) exploiting redundancy pertaining to word-vectors (intermediate encoder outputs) and eliminating the redundant vectors. b) determining which word-vectors to eliminate by developing a strategy for measuring their significance, based on the self-attention mechanism. c) learning how many word-vectors to eliminate by augmenting the BERT model and the loss function. Experiments on the standard GLUE benchmark shows that PoWER-BERT achieves up to 4.5x reduction in inference time over BERT with &lt;1% loss in accuracy. We show that PoWER-BERT offers significantly better trade-off between accuracy and inference time compared to prior methods. We demonstrate that our method attains up to 6.8x reduction in inference time with &lt;1% loss in accuracy when applied over ALBERT, a highly compressed version of BERT. The code for PoWER-BERT is publicly available at https://github.com/IBM/PoWER-BERT.

Motivation & Objective

• Motivate reducing BERT inference time without retraining or changing model size.
• Identify a new redundancy type: word-vectors (intermediate encoder outputs) that diffusion creates.
• Develop a dynamic, attention-guided method to select and retain word-vectors across encoders.
• Learn a retention configuration that specifies how many word-vectors to keep per encoder while preserving accuracy.

Proposed method

• Leverage diffusion of information in BERT’s self-attention to justify progressive word-vector elimination across encoders.
• Define a retention configuration (ell_1,...,ell_{12}) specifying how many word-vectors to retain at each encoder.
• Compute word-vector significance scores using attention: Sig_h(w)=sum_w' A_h[w',w], Sig(w)=sum_h Sig_h(w).
• Introduce a soft-extract layer that learns retention parameters r_j[k] (in [0,1]) to softly scale word-vectors by their sorted significance positions.
• Train in three steps: (1) fine-tune BERT; (2) configuration-search with soft-extract layers and lambda-regularized loss to derive mass and retention; (3) replace with hard extract and re-train.
• Query the loss: L(Θ,r) + λ * sum_j j * mass(j;r), where mass(j;r)=sum_k r_j[k], to obtain a monotone retention configuration.

Experimental results

Research questions

• RQ1 Can word-vectors, as opposed to whole encoders, be safely eliminated to speed up inference without hurting accuracy?
• RQ2 Does attention-derived significance accurately identify word-vectors whose removal minimally impacts predictions?
• RQ3 Can a learned retention configuration adaptively determine how many word-vectors to keep per encoder across inputs?
• RQ4 How does PoWER-BERT compare to encoder-pruning baselines and to compressed models (e.g., ALBERT) in accuracy-time trade-offs?

Key findings

• PoWER-BERT achieves up to 4.5x faster inference than BERT BASE with <1% accuracy loss on GLUE tasks.
• On ALBERT, PoWER-BERT yields up to 6.8x speedup with <1% accuracy loss on GLUE tasks.
• Dynamic, attention-based word-vector selection (Attn-WS) outperforms static (Head-WS, Rand-WS) strategies, especially on longer inputs.
• The learned retention configuration progressively removes word-vectors across encoders, while preserving the CLS-based decision pathway (CLS is not eliminated).
• Compared to DistilBERT, BERT-PKD, and Head-Prune, PoWER-BERT offers superior Pareto frontiers in accuracy vs. inference time across multiple GLUE datasets.
• Code is publicly available, enabling reproducibility and integration with production pipelines.