Reducing Transformer Depth on Demand with Structured Dropout

TL;DR

This paper introduces LayerDrop, a structured dropout method that enables dynamic depth adjustment in transformer models, improving efficiency and performance across multiple NLP tasks without additional fine-tuning.

Contribution

The paper presents LayerDrop, a novel structured dropout technique that allows for selecting sub-networks of varying depth from a single trained transformer, enhancing efficiency and performance.

Findings

Improved state-of-the-art results on multiple NLP benchmarks.

Enables efficient inference with variable model depth.

Produces higher quality small models compared to training from scratch or distillation.

Abstract

Overparameterized transformer networks have obtained state of the art results in various natural language processing tasks, such as machine translation, language modeling, and question answering. These models contain hundreds of millions of parameters, necessitating a large amount of computation and making them prone to overfitting. In this work, we explore LayerDrop, a form of structured dropout, which has a regularization effect during training and allows for efficient pruning at inference time. In particular, we show that it is possible to select sub-networks of any depth from one large network without having to finetune them and with limited impact on performance. We demonstrate the effectiveness of our approach by improving the state of the art on machine translation, language modeling, summarization, question answering, and language understanding benchmarks. Moreover, we show that our approach leads to small BERT-like models of higher quality compared to training from scratch or using distillation.