摘要
本文提出了输出特性具有多值和多阈值特性的神经元。这种新神经元的功能在于它用多个阈值把输入空间划分成多个超平带，并可把各个超平带按需求归类或赋以不同的输出值作为标号。以这类神经元为基础讨论了在解决奇偶校验问题和实现特定逻辑功能方面的优越性。理论上只需一个神经元就能实现任意位的奇偶校验或任意的逻辑功能，因而特别适合在模式识别和分类中应用。论文还研究了多阈值神经元的组合问题，发现二个多阈值神经元的特定组合可以发生阈值数量的倍增现象，这为复杂多阈值神经元的模拟计算与硬件实现提供了方便途径。
ANNs for pattern recognition with multi-valued(MV) and multi-thresholded(MT) transfer function as neurons' nonlinearities have been proposed. The function of multithresholded neurons (MTNs) lay in the fact that it partitioned the input space into several hyperplanar zones by using many thresholds in neurons' output transfer function, it also marked these different zones with different output values as required. On the basis of these neurons, the advantage of solving parity problem and specific logic problem was investigated. It had been shown that only one neuron was needed to solve N-bit parity problem or to perform arbitrary logic function.Therefore, it was especially suitable for pattern recognition and classification. The combination of MTNs was also investigated. It was found that two MTNs may result in the multiplication of threshold number of combined net.
出处
《电子学报》
EI
CAS
CSCD
北大核心
1996年第5期1-6,共6页
Acta Electronica Sinica
基金
国家863
八五攻关及军事电子预研基金
关键词
神经网络
多值
多阈值
模式识别
神经元
组合
Neural networks
Multi-valued
Multi-thresholded
Pattern recognition
Artificial neurons
