Adaptive noise schedule for denoising autoencoder

被引：19

作者：

Chandra, B. ^{[1
]}

Sharma, Rajesh Kumar ^{[1
]}

机构：

[1] Department of Mathematics, Indian Institute of Technology Delhi, New Delhi, India

来源：

Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) | 2014年 / 8834卷

关键词：

Deep learning - Neurons - Signal encoding;

D O I：

10.1007/978-3-319-12637-1_67

中图分类号：

学科分类号：

摘要：

The paper proposes an Adaptive Stacked Denoising Autoencoder (ASDA) to overcome the limitations of Stacked Denoising Autoencoder (SDA) [6] in which noise level is kept fixed during the training phase of the autoencoder. In ASDA, annealing schedule is applied on noise where the average noise level of input neurons is kept high during initial training phase and noise is slowly reduced as the training proceeds. The noise level of each input neuron is computed based on the weights connecting the input neuron to the hidden layer while keeping the average noise level of input layer to be same as that computed by annealing schedule. This enables the denoising autoencoder to learn the input manifold in greater details. As evident from results, ASDA gives better classification accuracy compared to SDA on variants of MNIST dataset [3]. © Springer International Publishing Switzerland 2014.

引用

页码：535 / 542