Applying BinaryWeights in Neural Networks for Sequential Text Recognition

With the development of deep learning, researchers have achieved lots of breakthroughs in many classical problems. Unfortunately, these progresses are demanding on the hardwares, especially GPU, causing a huge energy consumption. Therefore, how to implement these neural networks with lower requirements of hardwares is holding more and more attention. In this paper, two aspects of the work were done. Firstly, we build a deep learning framework that supports training and prediction with binarized neural networks. This framework include binarized layers, e.g. Convolution (Conv) and LSTM layers. It is based on our analysis of how to implement a binarized layer and train BianryWeights with it. Secondly, we construct a network with binarized layers which are implemented in our framework to achieve good performance on sequential text recognition. We also modify the network architecture in order to obtain the experimental results with little loss on accuracy.