Research On Synchronization Of Multiple Neural Networks

Synchronization is a common phenomenon in the nature, including the filed of theneural networks. It is of great potential significance when the synchronization of neuralnetworks is applied in artificial intelligence and other application domains. Nowadays,it is mainly the synchronization of two neural networks. At the beginning, the weightsare vectors at random. During the course of leaning, each of neural network receives thesame input vectors, calculate its outputs respectively and then exchange it. Based on this,each further adjusts its weight by a certain learning rule. After several limited rounds oflearning, the discrete weight of the neural networks will reach to the stage ofsynchronization and remain it in the further learning.The above mentioned characteristic of the neural network can be put into practice,such as the area of information safety. In other words, the synchronization of neuralnetworks can be used for the exchange of cryptographic key at the public channel. Thethesis is focused on the synchronization of multiple neural networks and its application.Firstly, it put forward several patterns of the synchronization of neural networks, such ascenter-learning pattern, distributed pattern, neighbor-learning pattern, majority-learningpattern and mixed synchronization pattern. Secondly, it analyzes the timing and effectsof synchronization of the different patterns using different learning rules, supported bylarge sum of experimental data. Lastly, it raises the feasibility of the application of thesynchronization of multiple neural networks in the areas of group key managementprotocol (GMKP) and identification verification. The thesis points out a new directionfor the research and application of synchronization of neural networks in the future.