BERT Loses Patience: Fast and Robust Inference with Early Exit

TL;DR

This paper introduces a patience-based early exit method for pretrained language models that enhances inference speed and robustness by dynamically stopping inference when intermediate predictions stabilize, improving efficiency and accuracy.

Contribution

It presents a novel plug-and-play early exit technique that couples internal classifiers with each layer of a PLM, improving speed and robustness over existing methods.

Findings

Improves inference efficiency by reducing the number of layers used.

Enhances model robustness and accuracy by preventing overthinking.

Achieves better accuracy-speed trade-off compared to existing early exit methods.

Abstract

In this paper, we propose Patience-based Early Exit, a straightforward yet effective inference method that can be used as a plug-and-play technique to simultaneously improve the efficiency and robustness of a pretrained language model (PLM). To achieve this, our approach couples an internal-classifier with each layer of a PLM and dynamically stops inference when the intermediate predictions of the internal classifiers remain unchanged for a pre-defined number of steps. Our approach improves inference efficiency as it allows the model to make a prediction with fewer layers. Meanwhile, experimental results with an ALBERT model show that our method can improve the accuracy and robustness of the model by preventing it from overthinking and exploiting multiple classifiers for prediction, yielding a better accuracy-speed trade-off compared to existing early exit methods.