忆阻神经形态系统在模式识别中的应用

发布时间：2018-05-09 15:19

本文选题：忆阻器 + 神经形态系统　；参考：《西南大学》2017年硕士论文

【摘要】：神经形态系统是仿人工神经网络构建的硬件系统,可以实现更高的信息处理和容错能力,被广泛应用于模式识别、机器学习、信号处理、图像处理等领域。忆阻器作为一个纳米级元件,具有非易失性/易失性、记忆性、可塑性、低功耗等特点,可以作为一个天然的突触。基于忆阻器的交叉架构则可以作为神经形态系统中天然的权值矩阵。由于不同的忆阻器具有不同的特性,为满足不同的需求,基于各类忆阻器设计的神经形态系统也越来越丰富。但这些系统多数存在三个问题,其一,通过传统工具对测试样本进行了大量的预处理;其二,网络的训练过程通常是通过线下系统实现的,而测试过程则是在神经形态电路系统上实现。其三,大多神经形态系统仿照传统数字系统的处理方法,而忽略了人脑的独特特性,比如遗忘特性。针对以上问题,本文构建了一种基于方差相关学习算法的遗忘忆阻神经形态系统,并将该系统成功应用于模式识别。在进行有效的手写数字图像识别之外,还研究了遗忘速率与识别效率之间的关系。本文的具体研究内容和成果如下:1、对经典惠普忆阻器和遗忘忆阻器的内部机制进行了阐述说明,并对其数学模型进行了理论推导。在一维遗忘忆阻器模型的基础上,介绍了改进后的三维遗忘忆阻器模型,并给出了三维遗忘忆阻器模型的单、双极和单双可逆条件。通过建立SPICE仿真模型,对这三种忆阻器模型的内部特性和突触行为进行了详细的比较和分析。2、基于一维遗忘忆阻器模型设计了一种神经形态电路系统。该系统是包含自学习电路系统、训练电路系统及识别电路系统的多层集成系统,可以实现样本的在线训练和识别功能。针对系统不同层所实现功能的不同,给出了各个部分的电路原理设计和功能仿真。3、基于样本的群体特征和个体特征,提出方差相关学习算法,实现对忆阻交叉架构矩阵在线的训练。该方法可以有效简化样本的预处理工作,同时便于电路系统的实现。4、将神经形态系统应用于手写数字图像的模式训练及识别。通过仿真验证了系统的功能和有效性。另外,进一步研究了遗忘忆阻器的遗忘因子?对识别结果的影响,发现不同区间的遗忘因子对识别效果具有不同程度的影响。
[Abstract]:Neural Morphology system is a hardware system constructed by artificial neural network, which can achieve higher information processing and fault tolerance. It is widely used in pattern recognition, machine learning, signal processing, image processing and other fields. As a nanoscale device, the memory device has the characteristics of non-volatile / volatile, memory, plasticity and low power consumption, so it can be used as a natural synapse. The cross structure based on amnesia can be used as a natural weight matrix in neural morphological system. Because different amnesia devices have different characteristics, in order to meet different needs, neural morphological systems based on various kinds of amnesia devices are more and more abundant. However, most of these systems have three problems: first, a large number of test samples are preprocessed by traditional tools; second, the training process of network is usually realized by offline system. The test process is implemented on the neural morphological circuit system. Third, most neural morphological systems mimic the traditional digital systems, while ignoring the unique characteristics of the human brain, such as forgetting. In order to solve the above problems, a forgetfulness and amnesia neural morphological system based on variance correlation learning algorithm is constructed in this paper, and the system is successfully applied to pattern recognition. In addition to effective handwritten digital image recognition, the relationship between forgetting rate and recognition efficiency is also studied. The specific research contents and results of this paper are as follows: 1. The internal mechanisms of the classic Hewlett-Packard (HP) amnesia and the amnesia amnesia are explained, and its mathematical model is theoretically deduced. On the basis of one dimensional amnesia model, the improved three dimensional amnesia model is introduced, and the single, bipolar and single double reversible conditions of the three dimensional amnesia model are given. The internal characteristics and synaptic behavior of the three kinds of amnesia models are compared and analyzed in detail by establishing the SPICE simulation model. A neural morphological circuit system is designed based on the one-dimensional amnesia model. The system is a multi-layer integrated system including self-learning circuit system, training circuit system and recognition circuit system, which can realize on-line training and recognition of samples. Aiming at the different functions realized by different layers of the system, the circuit principle design and function simulation of each part are given. Based on the group and individual characteristics of the sample, the variance correlation learning algorithm is proposed. The online training of memory cross-architecture matrix is realized. This method can simplify the preprocessing of the sample effectively and is convenient for the realization of the circuit system. The neural morphological system is applied to the pattern training and recognition of handwritten digital image. The function and effectiveness of the system are verified by simulation. In addition, the forgetfulness factor of amnesia is further studied. It is found that the forgetting factors of different intervals have different effects on the recognition results.
【学位授予单位】：西南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN60;TP391.4

【参考文献】