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.