用于联想记忆的回归网络的权值设计

NEURAL NETWORKS TRAINED FOR ASSOCIATIVE MEMORY

本文提出了一种连续型前馈回归网络用于进行模拟值样本的联想记忆,网络的权值通过BP算法学习得到.用这种方法可以保证每个待记样本在可预先设计的凸域内是吸引的,并给出了一种球形吸引域的设计方法.为加深理解,文中给出了此种回归网与全连接动态网(Hopfield网)权值间的关系.

This paper presents a learning method for designing associative memories using recurrent feedforward neural networks which can also be implemented by fully interconnected recurrent neural networks attached on a linear output layer. This technique guarantees that desired memories are stored and are attractive over the prescribed domains. The learning method can be traced back to BP algorithm.