Research Of Deep Learning Based On Time Accumulation And Adaptive Resonance Mechanism

ART2is a kind of non-supervised neural network based on the AdaptiveResonance Theory, and due to such advantages as rapid response and real-time learningabilities, ART2has been widely used in real-time clustering problems. ART2network isset up based on human cognitive rules and has very strong ability for learning,especially for the learning of new category, so it has great value for research.Firstly, this paper makes a depth research on the operating mechanism of ART2network, and based on this we analyze the problems of pattern drifting, losing ofamplitude information and the difficulty in setting vigilance parameter which exists inART2. Then we propose methods to solve these problems respectively, and prove theeffectiveness of each improved method by the simulation experiment, which makes thecomprehensive performance of ART2network promoted. Secondly, the paper does adepth research on memory forgetting and transformation mechanism of the cognitivetheory, and applies it to ART2network. The paper defines the concept of comprehensivememory strength (CMS), using it to record the memory value of each mode in ART2network, and adjusts the value of CMS in the learning process of the network constantly.By setting a reasonable threshold of CMS can makes the ART2network has the abilityto detect the new category. Meanwhile, combining the discipline that people distinguishthings that the more familiar things the faster to be recalled, the paper proposes amethod of selecting winning mode based on the value of CMS. By using the modewhich has the largest value of CMS as the winning mode, the operation complexity ofART2network is reduced. Based on the studies above, the performance of ART2network is improved and the function of it is consummated.In order to deal with a large number of classification problems related to time, thepaper does a depth research on TANN network, and on the basis of previous studies, weconstruct a depth learning system based on TANN and ART2(TANN-ART2). It not onlyretains the original time information in the time-series data, solving the problem thattraditional neural network is difficult to deal with time-series data, but also it can findthe new coming category from a large number of continuous time-series data in thelearning process, that makes the network show a strong learning ability andself-correction capability. So TANN-ART2has great value for application and research.Finally, this paper applies the TANN-ART2depth learning system in the keywordssearching for testing, getting the correct rate of71.5%and error rate of56.1%respectively, and detecting several new emerging words which has a high frequencyfrom the set. So proving that the TANN-ART2depth learning system has a strongability for dealing with time-series classification and detecting the new category.