Heterologous Normalization
Batch Normalization has become a standard technique for training modern deep networks. However, its effectiveness diminishes when the batch size becomes smaller since the batch statistics estimation becomes inaccurate. This paper proposes Heterologous Normalization, which computes normalization's mean and standard deviation from different pixel sets to take advantage of different normalization methods. Specifically, it calculates the mean like Batch Normalization to maintain the advantage of Batch Normalization. Meanwhile, it enlarges the number of pixels from which the standard deviation is calculated, thus alleviating the problem caused by the small batch size. Experiments show that Heterologous Normalization surpasses or achieves comparable performance to existing homologous methods, with large or small batch sizes on various datasets.
PDF Abstract
CIFAR-10
CIFAR-100
