An Improved Artificial Neural Network Model

The artificial neural network research is closely related to the understanding and research of human brain structure. More and more attention has been paid to simulate the study, cognition, memory of the human brain through artificial neural network, etc.Based on previous work, an improved artificial neural network model-the pattern neuron network was designed. This network introduced the rough classification criterion, the memory-forgetting mechanism and the memory cycle mechanism to realize the full simulation of the human brain's study cognition and memory process. The structure of the network is dynamically changed, which makes it self-learned and adaptive. The pattern neuron network can conduct multi-level classification and rapid recognition; it can be used to solve related problems of machine vision recognition and pattern classification. The contributions of this dissertation are summarized as follows:1.The SOM and ART1 network were implemented. According to the experiments, it analyzed the disadvantages of the two networks. The ART1 network is easy to be disturbed by input order of the sample. Aimed to solve this problem, we proposed an optimization algorithm, and we designed and implemented it. The optimization algorithm use the idea of "exclusive-OR" to calculate the matching degree, reducing classify error and improved the network's recognition accuracy. In the binary characters classification experiment, this improved algorithm reduced the sensitivity of the input sample sequence and got better classification performance.2. In order to solve the SOM and ART1 network's disadvantages, the pattern neural network was designed. This network has the function of both fine classification and rough classification by setting two vigilance parameters, while the SOM and ART1 network could only achieve either of these classifications. The vigilance parameters can be used to filter the sample vector and weight vector while computing the matching degree. This network avoided the iterative training process of the SOM. This network got higher recognition classification accuracy than SOM and ART1 by using the principle of "exclusive-OR". This network's output-neurons can dynamic grow along with the increasing of the pattern. Compared the three networks mentioned above through experiment.3. The pattern network's structure and algorithm was optimized to get a better simulating of the brain's cognitive mechanism. Introduced the memory-forgetting mechanism to the pattern neuron network's structure; provide two methods to process different sample set, and this accelerate the recognition speed; provided different rough classification criteria to enhance network's flexibility; reduced the network computing amount while process small sample set. The new network was compared with SOM and ART1 and the unimproved pattern neural network through experiment. In the experiment, the network's recongnition speed for small sample set is 1.95 faster than ART1, and is 58.13 faster than SOM; and is 1.55 faster than the unimproved pattern neural network; the network's speed for large sample set recongnition is 1.92 faster than ART1, and is 64.75 faster than SOM, and is 2.48 faster than the unimproved pattern neural network.