arrow
返回

DropDim: A Regularization Method for Transformer Networks

delete2022-01-01
delete4
delete
OA
AI
H
Hao Zhang
D
Dan Qu
K
Keji Shao
X
Xukui Yang *
DOI:10.1109/LSP.2022.3140693delete
delete原文链接
delete原文求助
delete分享
delete收藏
摘要

摘要

En 中文
We introduce DropDim, a structured dropout method designed for regularizing the self-attention mechanism, which is a key component of the transformer. In contrast to the general dropout method, which randomly drops neurons, DropDim drops part of the embedding dimensions. In this way, the semantic information can be completely discarded. Thus, the excessive co-adapting between different embedding dimensions can be broken, and the self-attention is forced to encode meaningful features with a certain number of embedding dimensions erased. Experiments on a wide range of tasks executed on the MUST-C English-Germany dataset show that DropDim can effectively improve model performance, reduce over-fitting, and show complementary effects with other regularization methods. When combined with label smoothing, the WER can be reduced from 19.1% to 15.1% on the ASR task, and the BLEU value can be increased from 26.90 to 28.38 on the MT task. On the ST task, the model can reach a BLEU score of 22.99, an increase by 1.86 BLEU points compared to the strong baseline.
Keyword:
Task analysis
Transformers
Smoothing methods
Semantics
Training
Neurons
Decoding
End-to-end
transformer
regularization
dropout

期刊

IEEE Signal Processing Magazine 封面图
IEEE Signal Processing Magazine
IF:
9.6
论文数:
1.1W
被引数:
1.7W

机构

暂无机构信息
引用论文

引用论文

err分享
err收藏
Machine Learning in Materials Science
err2021-08-24
err0
PREAI
errKeith T. Butler; Felipe Oviedo; Pieremanuele Canepa
err分享
err收藏
Sucrose Preference and Novelty-Induced Hypophagia Tests in Rats using an Automated Food Intake Monitoring System使用自动化食物摄入监测系统进行的蔗糖偏好和 novelty-induced hypophagia 测试
err2020-05-08
err0
PREAI
errMartha Anna Schalla; Stephanie Gladys Kühne; Tiemo Friedrich; Vivien Hanel; Peter Kobelt; Miriam Goebel-Stengel; Matthias Rose; Andreas Stengel
err分享
err收藏
学者 查看更多内容