在原有工作基础上，利用所提出的数字型稀疏RAM神经元代替RAM式及常规型神经元，设计出一个易于硬件实现的自适应模式识别系统模型，使其能处理大维数输入模式和大样本集类的实时性识别问题，对此模型开展其学习能力，泛化能力，识别能力，函数逼近能力及对CMAC模型的推广能力的研究。从而为大样本集的自适应模式识别技术提供一个新的实现方法。

Proposed a sparse random access memory (RAM) neuron concept, constructed a sparse-RAM-based system model for parallel pattern recognition, designed and implemented its fast learning algorithm. The model overcomes the saturation problem of the existing models such as WISARD (Wil.-Igor-Stonham Adaptive Recognition Device) and reduces the cost for implementation. Based on these results, further presented a general unified model and learning algorithm for memory-based neural networks. The model formulates Kanerva's sparse distributed memory (SDM), Albus' cerebellar model articulation controller (CMAC)，Alexsander's WISARD and Tattersall's single layer look-up perceptron into a common framework, shown convergence of its learning algorithm, examined its recognizing feasibility on ORL face database and good approximate ability to any continuous functions, pointed out the influence of both n-tuple size and their distribution on the recognition and approximation performances, come up with and designed a boosting n-tuple method, which not only optimizes the system performance, but also extends its learning adaptation to sub-pattern selection, in this way, a new method for feature selection is put forward. In addition, with the support vector machine, an effective recognition system for computer user's keystroke patterns was realized such that the security of access to computer is enhanced. Furthermore, by combining the fuzzy technique and the subspace method, we presented a fuzzy subspace recognition method which raises the recognition rate and reliability.