Based on the Tensorflow Framework HCMAC Neural Network Design

• Main Authors: Guan-Min Wei, 魏冠民
• Other Authors: Yung-Feng Lu
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/t8m2vy

碩士 === 國立臺中科技大學 === 資訊工程系碩士班 === 106 === ABSTRACT Deep neural networks require high performance computing and operations.Among them, the machine learning framework TensorFlow is a neural network implementation tool.The calculations expressed using TensorFlow can be flexibly applied to various systems.Paspberry Pi, Android, Windows, iOS, Linux to server,With little or no modification, portability ranges from portable devices to large distributed systems. Even tens of millions of computing devices.The entire system is very flexible,Mainly explore very large deep neural networks,Data flow graph deep learning provides convenience.And assume that many subgraphs contain distributed execution in this dataflow graph,These executions are manually defined through the step call API. CMAC can be regarded as a mapping method. It has the characteristics of basic functional network (BFN) and supervised learning. It has fast learning convergence and approximation effects, and has the characteristics of simple structure and good local generalization effect. However, since the hyper-cubic blocks of CMAC are evenly distributed, the values of the output in each partition are equal, and in order to solve the problem of huge storage in high-dimensional problems, The performance that causes it depends to a large extent on input space quantization. In order to solve the above two problems, the Gaussian function is added to the hypercube block to form the non-fixed variable micro-partition block, and the HCMAC is cited on the high-dimensional classification problem, which has the ability of fast learning and low memory requirements, and the traditional Compared with CMAC, it is a good neural network in solving high-dimensional classification problems. In this paper, the simulation results show that HCMAC applied to Tensorflow has good approximation ability and convergence performance.