基于竞争的联想存储器学习算法

提出两种基于竞争的神经网络联想存储器学习算法—CC算法和ACC算法 ,并证明算法得到的神经网络对任一输入模式的竞争收敛性 ,由CC算法得到的网络 ,利用 p n个神经元存储p个n维样本模式 ;每个样本点都是吸引中心 ,不存在假吸引中心 ;对任一输入模式 ,总被吸引到与之海明距离最小的样本点上 ;不产生拒识点 .ACC算法是CC算法的改进形式 ,所得网络可在自适应学习中收敛 ,竞争次数较CC算法大大降低 本文算法得到的网络在存储容量、容错能力方面好于Hopfield联想存储器及作为联想存储器使用的BP网络 .

TWO KINDS OF ASSOCIATIVE MEMORIES BASED ON COMPETITION

Two new learning algorithms called CC and ACC for associative memory are presented in this paper.The algorithms employ the competing-classifying idea with the competition convergency for any input pattern.CC neural networks use p n neurons to save p samples of n dimension,any of which becomes an attractive point and no additional attractive points are generated.For any input pattern,the CC network always outputs a sample pattern with the least Hamming distance to the input,which leads to a good error-torlerance property.ACC algorithm is improved from CC algorithm.By the adaptive learning of the neural networks,the competing times are made much less than that of CC neural networks.The neural networks comstructed by the algorithms of this paper are better than HAM and BP networks used as associative memory in consideration of capacity and error-tolerance property.