Research On Kernel-based Enhanced Associative Memory Models With Applications

Associative memory (AM) is one of the most important functions of human brain. It is the source of logic thinking and image thinking, deduction and innovation. Thus, the trial to have the existing machine mimic some simple functions of brain is always the human beings'goal. Through more than twenty years'research, associative neural networks have been developed into an important class of artificial neural networks. And they have been widely applied to many fields such as optimizing computation, error correction coding, data compression, and pattern recognition.In this dissertation, we improve on AM models closely following the two cores of the AM research, learning algorithm and network architecture. With the popular'kernel method'in machine learning and the extensively-existed'small-world networks'in society, we propose the corresponding imrpoved AM models, and apply them to face recogntion and handwriting recogntion. The main contributions of this dissertation are summerized as follows:Firstly, Fuzzy Morphological Associative Memories (FMAMs) are proposed. The properties of FMAMs are greatly different from those of traditional AMs and Fuzzy Associative Memories (FAMs). Similar to MAMs, they have the good tolerance to single erosive or dilative noise. Auto-FMAMs have more attractive advantages such as unlimited storage capacity, perfect recall guarantee, and one-shot recall speed. The theoretical proofs about the above properties have also been given. Besides, FMAMs can also be explained using the terms in FAMs. The comparison with FAMs implies that under certain conditions, FMAMs can be viewed as another new encoding way of fuzzy rules.Secondly, with the widely-used kernel method, we overcome FMAM's shortcoming of the extreme vulnerability to noise of mixing erosion and dilation, and present the Enhanced FMAM (EFMAM) based on the empirical kernel map. The essence of EFMAM is that, without destroying the original AM's structure and encoding strategies, the patterns in the original space is mapped into the similarity space, which is spanned by the training patterns, by means of empirical kernel map. Thus the double noise tolerance is achieved. One is due to the empirical kernel map and the other due to the auto-FMAMs. EFMAM not only inherits the good properties of auto-FMAMs but also solves the hybrid-noise problem exploiting the new...