Research Of Application Of ART2 Network With Memory Strength

ART2 is a kind of self-organizing neural network which is based on adaptive resonance theory. It carries out the classification by using competitive learning and self-steady mechanism, and can learn by itself in dynamic environment with noise and without supervision. Its learning process can recognize learned models fast and be adapted to new unknown objects rapidly. This kind of neural network can simulate human's brain to recognize things.When the traditional ART2 neural network recognizes input patterns, there is a problem of pattern drifting existed. In order to solve that issue, this paper gives out an algorithms of ART2 based on slow weight update rule, which can slow down the learning rate, reduce the speed of pattern drifting in classification for input mode, and can classify objects rightly. The result of experiment shows that to a certain extent the improved neural network solves the problem of pattern drifting.In order to imitate the course of learning and cognizing things with the human brain better, this article introduced the human brain's Memorizing-forgetting mechanism to the ART2 Network, then used the Memory Strength as the basis of sequence for recognition with existing patterns, thus this network model would be improved better. Through simulation for recognition and classification with experimental samples, we prove that the ART2 network with Memorizing-forgetting mechanism could recognize experimental samples with less time in recognition than original ART2 network, and improve the efficiency of network.This paper still applied the ART2 neural network with memory strength into face recognition. After the dimensions of face images in ORL database are reduced by using the method of PCA or wavelet transform, the face vectors which we got are samples which will be input ART2 neural network. Through simulation for recognition and classification with ORL face database, we prove that the recognition accuracy of ART2 neural network that uses the recognition criteria of the angle's cosine is higher than the recognition accuracy of other methods that uses the recognition criteria of Euclidean distance or block distance, and that the improved ART2 neural network could save more recognition time than original ART2 neural network, especially for the large amounts of data the efficiency can be increased obviously. The features of cognitive learning function of ART2 neural network with the human brain's memorizing-forgetting mechanism are the foundation for its application in the actual environment with large mounts of samples and many patterns.