The Research On Improving Of Architecture And Trainning Performance Of Deep Neural Network

Deep learning technology has been applied into various of business, it would be used to offer the more intelligentized information. By combining polytype neural network layer, deep neural network module will be made up, and compute the raw input data, use Backpropagation algorithm to learn the more abstract representation of raw data, and then improve the accuracy of tasks for object detection or image classification. The key point of deep learning technology for classification to improve accuracy is that how deep is the deep neural network, how plentiful of the type of neural network’s layer, whether it is reasonable to the architecture of network, and if the trainning efficiency is acceptable.To improve the accuracy of task for image classification using deep neural network, we have present a new way of building neural network with different architecture and a new method to improve the trainning efficiency in this paper, and verified them.In the aspect of designing an architecture of network, we present a deep neural network based on multiple feature extraction network module, by separating the layers of stage used for feature abstraction to multiple traces, and compute the same trainning samples on each traces with no interference between, then we use backpropagation algorithm to fine-turning the parameters within the network moudle separately, learn the feature of raw images, and output a probability distribution according to classes of all images one for each trace. In order to enrich the method of feature extraction, we also designed a new feature extraction netwokr moudle based on autoencoder in this paper, it constitutes our multi-trace feature extraction network module along with feature extraction stage of Convolutional Neural Network. After stage of feature extraction, there will be multiple group of predicting probability distribution for image classification, they will be input to the network for error analysis designed in this paper, convert these data and then compute to output a new predicting probability distribution using our function, then it will be more close to the ground of truth.In the aspect of improve the trainning efficiency for our multi-trace network, in this paper, we put the redundant intermediate result, which is produced when trainning in progress, to specific disk space, this will avoid ruleless data switch between memory and disk. Then we improve the efficiency of system’s I/O scheduling algorithm by introducing the theory of Genetic Algorithm, maintain the key parameters of I/O scheduling algorithm independently, and made it could be adapt to I/O pressure from upper-layer application, then keep the best combination of key parameters, to improve the capacity of I/O requests scheduling. Meanwhile, in this paper, we also have designed and built a reliable operating system, which can be used to guarantee facticity of our I/O pressure data measured during experiment on trainning our mutil-trace network.