🤖 AI Summary
Existing BERT early-exit methods rely solely on per-sample local signals (e.g., entropy) to determine exit decisions, neglecting inter-class global structure—leading to biased reliability estimation and suboptimal exit choices.
Method: This work introduces prototype networks into early-exit mechanisms for the first time, proposing a distance-enhanced reliability assessment paradigm that jointly models local entropy and Euclidean distance to class prototypes. A dual-signal, synergistic hybrid gating strategy is designed and integrated into BERT’s hierarchical inference architecture—requiring zero additional parameters or computational overhead.
Contribution/Results: The method achieves significant improvements over state-of-the-art approaches on the GLUE benchmark across multiple acceleration ratios, consistently attaining higher accuracy. It exhibits strong generalization across diverse tasks and datasets, while offering enhanced interpretability through geometrically grounded exit decisions based on prototype distances and uncertainty.
📝 Abstract
Early exiting has demonstrated its effectiveness in accelerating the inference of pre-trained language models like BERT by dynamically adjusting the number of layers executed. However, most existing early exiting methods only consider local information from an individual test sample to determine their exiting indicators, failing to leverage the global information offered by sample population. This leads to suboptimal estimation of prediction correctness, resulting in erroneous exiting decisions. To bridge the gap, we explore the necessity of effectively combining both local and global information to ensure reliable early exiting during inference. Purposefully, we leverage prototypical networks to learn class prototypes and devise a distance metric between samples and class prototypes. This enables us to utilize global information for estimating the correctness of early predictions. On this basis, we propose a novel Distance-Enhanced Early Exiting framework for BERT (DE$^3$-BERT). DE$^3$-BERT implements a hybrid exiting strategy that supplements classic entropy-based local information with distance-based global information to enhance the estimation of prediction correctness for more reliable early exiting decisions. Extensive experiments on the GLUE benchmark demonstrate that DE$^3$-BERT consistently outperforms state-of-the-art models under different speed-up ratios with minimal storage or computational overhead, yielding a better trade-off between model performance and inference efficiency. Additionally, an in-depth analysis further validates the generality and interpretability of our method.