CascadeBERT: Accelerating Inference of Pre-trained Language Models via Calibrated Complete Models Cascade

TL;DR

CascadeBERT addresses bottlenecks in dynamic early exiting by replacing single-model early exits with a cascade of complete models and a difficulty-aware calibration. The approach uses sequential evaluation from small to large models, exiting when confidence exceeds a threshold, complemented by a regularization term that aligns predictions with instance difficulty. Empirical results on GLUE show substantial speed-ups with improved accuracy and calibration over baselines, and analyses confirm more discriminative representations and generalizability to RoBERTa. This framework offers a practical, calibrated path to efficient deployment of large pretrained language models across diverse tasks.

Abstract

Dynamic early exiting aims to accelerate the inference of pre-trained language models (PLMs) by emitting predictions in internal layers without passing through the entire model. In this paper, we empirically analyze the working mechanism of dynamic early exiting and find that it faces a performance bottleneck under high speed-up ratios. On one hand, the PLMs' representations in shallow layers lack high-level semantic information and thus are not sufficient for accurate predictions. On the other hand, the exiting decisions made by internal classifiers are unreliable, leading to wrongly emitted early predictions. We instead propose a new framework for accelerating the inference of PLMs, CascadeBERT, which dynamically selects proper-sized and complete models in a cascading manner, providing comprehensive representations for predictions. We further devise a difficulty-aware objective, encouraging the model to output the class probability that reflects the real difficulty of each instance for a more reliable cascading mechanism. Experimental results show that CascadeBERT can achieve an overall 15\% improvement under 4$\times$ speed-up compared with existing dynamic early exiting methods on six classification tasks, yielding more calibrated and accurate predictions.

Figures (5)

• Figure 1: An easy instance with a large word overlap (colored in orange) between the premise and the hypothesis from the MNLI dataset. The classifiers in shallow layers of a dynamic early exiting model cannot predict correctly, while BERT-Complete turc2019well, a small BERT pre-trained from scratch with the same size can make a correct and confident prediction.
• Figure 2: Performance comparison utilizing different models with the same number of layers on MNLI-m and SST-2. Complete models capable of extracting semantic-level information clearly outperform models like DeeBERT which overlooks the high-level semantic features.
• Figure 3: DIS (%, higher is better, see Eq. \ref{['eq:dis']} in Section \ref{['subsec:dis']}) heatmap of different models on the development set of MNLI and SST-2. The DIS of internal off-ramps in the DeeBERT of shallow layers is lower than that of BERT-$k$l and BERT-Complete, which leads to more wrongly emitted instances. The exiting decisions in shallow layers of DeeBERT thus can be unreliable.
• Figure 4: t-SNE visualization of instance representations of different class in DeeBERT and our CascadeBERT at the second layer. The instance representations of our CascadeBERT exhibit a more distinct boundary between different classes, helping the following classifier to make accurate predictions. Best viewed in color.
• Figure 5: Task performance on the validation set and speed-up ratio trade-off curve comparison of a 2-model cascade (orange square) and a 3-model cascade (blue circle) on SST-2 and MNLI-m.