论文笔记（一）基于轮流训练的松弛卷积神经网络对手写字附体的识别（MINIST和ICDAR’13 Competition 汉字库）

Handwritten Character Recognition by Alternately Trained Relaxation Convolutional Neural Network  

Published in:

Frontiers in Handwriting Recognition (ICFHR), 2014 14th International Conference on( 论文下载)

 The differences between traditional CNN and  proposed RCNN and ATR-CNN are as follows:

 (1) R-CNN consists of relaxation convolution layers whose neurons within a feature map do not share the same convolutional kernel  as convolutional layers; (2) ATR-CNN further adopts  alternate training strategy, i.e., the weights parameters of a  certain layer do not change by back-propagation algorithm  given a training epoch. Relaxation convolution can be considered to enhance the learning ability of the neural network,  while alternate training is to regularize the network during  training because relaxation connections sharply increase the  number of parameters.