Current browse context:
cs.NE
Change to browse by:
References & Citations
Computer Science > Neural and Evolutionary Computing
Title: Scale Normalization
(Submitted on 26 Apr 2016)
Abstract: One of the difficulties of training deep neural networks is caused by improper scaling between layers. Scaling issues introduce exploding / gradient problems, and have typically been addressed by careful scale-preserving initialization. We investigate the value of preserving scale, or isometry, beyond the initial weights. We propose two methods of maintaing isometry, one exact and one stochastic. Preliminary experiments show that for both determinant and scale-normalization effectively speeds up learning. Results suggest that isometry is important in the beginning of learning, and maintaining it leads to faster learning.
Link back to: arXiv, form interface, contact.