Normalized Convolutional Neural Network
Authors:
Dongsuk Kim,
Geonhee Lee,
Myungjae Lee,
Shin Uk Kang,
Dongmin Kim
Abstract:
In this paper, we propose Normalized Convolutional Neural Network(NCNN). NCNN is more fitted to a convolutional operator than other nomralizaiton methods. The normalized process is similar to a normalization methods, but NCNN is more adapative to sliced-inputs and corresponding the convolutional kernel. Therefor NCNN can be targeted to micro-batch training. Normalizaing of NC is conducted during c…
▽ More
In this paper, we propose Normalized Convolutional Neural Network(NCNN). NCNN is more fitted to a convolutional operator than other nomralizaiton methods. The normalized process is similar to a normalization methods, but NCNN is more adapative to sliced-inputs and corresponding the convolutional kernel. Therefor NCNN can be targeted to micro-batch training. Normalizaing of NC is conducted during convolutional process. In short, NC process is not usual normalization and can not be realized in deep learning framework optimizing standard convolution process. Hence we named this method 'Normalized Convolution'. As a result, NC process has universal property which means NC can be applied to any AI tasks involving convolution neural layer . Since NC don't need other normalization layer, NCNN looks like convolutional version of Self Normalizing Network.(SNN). Among micro-batch trainings, NCNN outperforms other batch-independent normalization methods. NCNN archives these superiority by standardizing rows of im2col matrix of inputs, which theoretically smooths the gradient of loss. The code need to manipulate standard convolution neural networks step by step. The code is available : https://github.com/kimdongsuk1/ NormalizedCNN.
△ Less
Submitted 18 May, 2020; v1 submitted 11 May, 2020;
originally announced May 2020.
