The Research Of Information Processing Based Onneural Network

This paper first introduces neural network, and then focuses on the application of neural network in comprehensive evaluation and handwritten digit recognition, and finally make a summary and outlook.Comprehensive evaluation has been widely used in social, scientific, management and other fields, and there are many ways of comprehensive evaluation, including fuzzy comprehensive evaluation method, hierarchical analysis, expert evaluation method, etc. However, in comprehensive evaluation, it is difficult to determine the weights and membership, which are the two major difficulties. In comprehensive evaluation, many ways have a strong subjective element, which affect the accuracy of the final evaluation. To get more objective and reasonable weights and membership, and improve the accuracy of the final evaluation, we use fuzzy neural network to adjust the weights and membership of comprehensive evaluation. This kind of neural network is used to evaluate the dynamic nature of the relay, and the results are objective and reasonable.Handwritten numeral recognition is widely used in many fields. In handwritten numeral recognition there are many recognition ways. Because BP neural network can approximate any nonlinear mapping, and has good adaptability, self-organization, etc., so it is widely used in handwritten numeral recognition. But there are some defects such as slow convergence speed and it is easy to fall into the "local minima" and others, so scholars have done lots of improvement, the VLBP improved algorithm is one kind of common improvement. However, there exists low efficiency and not smooth convergence in VLBP algorithm. To solve this problem, this paper has improved the learning rate BP algorithm through building up the functional relationship between the error E and the learning rateη, this improved BP algorithm adopts continuous dynamic adaptive learning rate. Compared with VLBP, the simulation results show this improved variable learning rate BP algorithm quickens up the convergence speed, and makes the error curve relatively stable from start to finish. After this improved learning rate BP algorithm is applied to handwritten digit recognition, the recognition effect is better.Sigmoid is used as activation function in standard BP algorithm commonly, but it is easy to make neural network fall into local minima. So this article makes some improvement on the sigmoid activation function. First we judge which samples exist okp≈0 or (1-okp)≈0, and then change the Okp which is closer to 0 to to tkp-Okp, or change the Okp which is closer to 1 to 1-(tkp-Okp) through directly modifying the parameter value, finally through this modify the local minima can be avoided. In the handwritten numeral recognition process, the individual characteristics has low recognition rate, so we propose one kind of nine-segment projection features, and combine it with the point structure. Thus one kind of new combination features is input to improved neural network. Because the combination characteristics have the advantages of statistic features and structure features, so the recognition rate raises a lot obviously.