The Research On The Phase Transition And The Anti-noise Performance In The DeSTIN

The Phase transition is widespread in nature; it exists in many problems of the Computer Science and the artificial intelligence research fields. And the phase transition has much to do with the structure of the problem itself. Also the research on the phase transition contributes to the improvement of the problem solving. The meaning of the study of phase transition is, therefore, its results to design effective have guiding significance for the tools and methods of solving problem.In recent years, deep learning network is all attention; they in different areas are showing surprising ability. However, how to improve the Connectionism calculation model under the condition of high noise performance, there are still a lot of work to be done.According to its reaction to different types of noise, this paper conducted on the application of phase transition and its research work. Studying the phase transition of the DeSTIN is to optimize the network itself, improve its robustness, and expand its application field.The contribution of research can be summarized as the following three points:1、According to the reasons got through analysis, this paper points out when the belief value variation of the clustering center exceed the coefficient of variation k, the recognition rate of the DeSTIN will generate the phase transition.2、This paper proofs the correctness of the above reasoning from the two groups of different experiment.3、According to the reasons got through analysis, this paper provides two methods to the noise immunity of the DeSTIN. The first is to use the traditional noise filtering to optimize and improve the network clustering operating. The second method is to use the transfer learning ideology to optimize and improve the network clustering operating of testing database utilizing existing database knowledge, through the depth digging to the feature of the image gotten already, that is the output of the network training.Having compared with two optimizing methods, the paper find the second one can avoid the happening of the phase transition and improve the robustness of the network to the noise. And at the same noise level, the recognition rate of the DeSTIN is higher than other deep learning networks’.The results of this paper will not only help reveal the phase transformation mechanism of a calculation connectionism model, but also it can improve the anti-noise ability of the network.