A dropout distribution model on deep networks

Fengqi Li; Helin Yang

doi:10.1117/12.2243971

29 August 2016 A dropout distribution model on deep networks

Fengqi Li, Helin Yang

Proceedings Volume 10033, Eighth International Conference on Digital Image Processing (ICDIP 2016); 1003360 (2016) https://doi.org/10.1117/12.2243971
Event: Eighth International Conference on Digital Image Processing (ICDIP 2016), 2016, Chengu, China

Abstract

Dropout is proved to have a good ability of controlling overfitting and improving deep networks’ generalization. However, dropout adopts a constant rate to train the parameters of each layer, reducing the classification accuracy and efficiency. Aiming at this problem, the paper proposes a dropout rate distribution model by analyzing the relationship between the dropout rate and the layers of the deep network. First, we gave the formal description of the dropout rate to reveal the relationship between the dropout rate and the layers of the deep network. Second, we proposed a distribution model for determining the dropout rate in each layer training. Experiments are performed on MNIST and CIFAR-10 datasets to evaluate the performance of the proposed model by comparison with networks of constant dropout rates. Experimental results demonstrate that our proposed model performs better than the conventional dropout in classification accuracy and efficiency.

Citation Download Citation

Fengqi Li and Helin Yang "A dropout distribution model on deep networks", Proc. SPIE 10033, Eighth International Conference on Digital Image Processing (ICDIP 2016), 1003360 (29 August 2016); https://doi.org/10.1117/12.2243971

ACCESS THE FULL ARTICLE

INSTITUTIONAL
Select your institution to access the SPIE Digital Library.

SELECT YOUR INSTITUTION

PERSONAL
Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.

PERSONAL SIGN IN

No SPIE Account? Create one

PURCHASE THIS CONTENT

SUBSCRIBE TO DIGITAL LIBRARY

50 downloads per 1-year subscription

Members: $195

Non-members: $335 ADD TO CART

25 downloads per 1 - year subscription

Members: $145

Non-members: $250 ADD TO CART

PURCHASE SINGLE ARTICLE

Includes PDF, HTML & Video, when available