Research On Associative Memory Neural Networks

In this paper, an unconventional associative memory (AM) and its hardware implementation have been researched.The research is aimed at the problems in current associative memory as follows:1. In order to memorize any new pattern, existing AM must do a large amount of computing to update its all synapse weights. It causes the working of searching for suitable weights difficult.2. AMs are mainly nonlinear dynamic models in themselves there are stability, robustness, and convergence having to be satisfied. Any AM should have large storage capacity, good fault tolerance, and few pseudo attractors. Obviously, it is difficult to satisfy both requirements.3. Although it will reduce the importance of any link for memory that every link is used to store each pattern, it cannot remove the impact of any link on each stored pattern. A few broken links could lead to AM losing all capability of memorizing.4. In an AM, there may be a number of stored patterns which all have the minimum Hamming distance with input pattern having not been stored, but it can only retrieve one of them.5. The input pattern for recalling must be completely determined, but incomplete input pattern is more practical, because one of the important functions of AM is to restore complete information by association according to incomplete input information.The significance of the research in this paper is by trying different methods to seek new type of AM that can overcome or has no the said above problems. The basic idea is summarized as follows:1. Every pattern is stored in one or more than one directional closed loop (named as PL--'pattern loop') which is composed of links and passes through every neuron of the AM only one time.2. Synapse weight is replaced by 'link state' and 'inhibited path'. The former represents the information of stored pattern, and the latter prevents signals from passing through PLs that store spurious patterns,...