DeeBERT: Dynamic Early Exiting for Accelerating BERT Inference
This addresses the bottleneck of real-time deployment for NLP applications, though it is incremental as it builds on existing BERT architectures.
The paper tackled the problem of slow inference in large pre-trained language models like BERT by proposing DeeBERT, a method that allows samples to exit early without full model processing, saving up to ~40% inference time with minimal quality degradation.
Large-scale pre-trained language models such as BERT have brought significant improvements to NLP applications. However, they are also notorious for being slow in inference, which makes them difficult to deploy in real-time applications. We propose a simple but effective method, DeeBERT, to accelerate BERT inference. Our approach allows samples to exit earlier without passing through the entire model. Experiments show that DeeBERT is able to save up to ~40% inference time with minimal degradation in model quality. Further analyses show different behaviors in the BERT transformer layers and also reveal their redundancy. Our work provides new ideas to efficiently apply deep transformer-based models to downstream tasks. Code is available at https://github.com/castorini/DeeBERT.