Intermediate pattern stimulus discrimination Neural Network modeling

• Main Authors: Chung-Han Tung, 董忠漢
• Other Authors: Sigmund Hsiao
• Format: Others
• Language: zh-TW
• Published: 2000
• Online Access: http://ndltd.ncl.edu.tw/handle/62198558423662504376

碩士 === 國立中正大學 === 心理學研究所 === 88 === The simulation and comparison of intermediate pattern categorization of human in four different neural networks was studied. After learning six pairs of original alphabet or numerical pattern stimuli, the performance of neural network in categorizing intermediate pattern stimuli was tested. In Exp 1, human subjects were tested using 6 pairs of alphabet or numerical intermediate patterns. In Exp 2 we used a generalized Hebbian learning neural network to learn original pattern stimuli and tested it by intermediate pattern stimuli. Results indicate that the performance of Hebbian learning neural network was similar to the generalization gradients in classical conditioning. In Exp 3 and Exp 4 we used perceptron, LMS and back-propagation to learn to discriminate original 6 pairs of pattern stimuli and tested it with categorization of intermediate stimuli of the learned pairs. The results of Exp 3 and Exp 4 were compared with Exp 1. We found that all human subjects, perceptron, LMS and back-propagation could learn to discriminate original patterns, but the performances of discriminating intermediate pattern stimuli were differentiable between these four by the intermediate pattern categorization task. Overall, the results indicate that using of specialized task might differentiate neural network performances that were originally undifferentiable, and the results were comparable with the performance of human subjects. The present approach might be used to study neural network simulation in human perception and categorization and drew inferences upon human cognitive processes.