Normalizing the Normalizers: Comparing and Extending Network Normalization Schemes

Mengye Ren^{* 1}, Renjie Liao^{* 1}, Raquel Urtasun¹, Fabian H. Sinz², Richard S. Zemel<sup>1,3

1</sup>Department of Computer Science, University of Toronto, Toronto ON, CANADA
²Department of Neuroscience, Baylor College of Medicine, Houston TX, USA
³Canadian Institute for Advanced Research, Toronto ON, CANADA
^*Equal contribution

Abstract

Normalization techniques have only recently begun to be exploited in supervised learning tasks. Batch normalization exploits mini-batch statistics to normalize the activations. This was shown to speed up training and result in better models. However its success has been very limited when dealing with recurrent neural networks. On the other hand, layer normalization normalizes the activations across all activities within a layer. This was shown to work well in the recurrent setting. In this paper we propose a unified view of normalization techniques, as forms of divisive normalization, which includes layer and batch normalization as special cases. Our second contribution is the finding that a small modification to these normalization schemes, in conjunction with a sparse regularizer on the activations, leads to significant benefits over standard normalization techniques. We demonstrate the effectiveness of our unified divisive normalization framework in the context of convolutional neural nets and recurrent neural networks, showing improvements over baselines in image classification, language modeling as well as super-resolution.

Full Paper

[pdf]

Poster

[pdf]

Code

[link]

Cite

@inproceedings{ren17norm,
  author    = {Mengye Ren and Renjie Liao and Raquel Urtasun and 
               Fabian H. Sinz and Richard S. Zemel},
  title     = {Normalizing the Normalizers: Comparing and Extending Network 
               Normalization Schemes},
  booktitle = {ICLR},
  year      = {2017}
}