Besearch Of The Design Of Online Learning Accelerator

SVM(Support Vector Machine) theory has been recently drawed much more attention as an important research branch in Machine Learning technologies. Its theory is becoming a hot spot in research of Machine Learning.One of the characteristics of Support Vector Machine is pursuiting the target of minimum structural venture. And it is this characteristic that makes Support Vector Machine theory be widely accepted by its application value. As Support Vector Machine developed from theory research to application in industry, it is required to have more advanced ability. One of them is supporting On-line learning ability.On-line Learning algorithm is also another important research branch in Machine Learning. Due to the excellent theory value and application value of Support Vector Machine, its on-line learning algorithm has become another hot spot in Machine Learning. Several kinds of software algorithm of on-line learning of Support Vector Machine have been proposed in the recent years, and these algorithm has received great effects in applications. As the embedded system develops, the hardware algorithm of on-line learning of Support Vector Machine is revealing its theorical value and applicational value. But very few hardware algorithm of on-line learning of Support Vector Machine is proposed these years and, at the same time, seldom software algorithm deserves the concept of "hardware friendly". As a result, there is great value and developing space in this research branch.In order to solve the above problems, the research in this paper is listed as follows:1. A brief introduction to develop background, application field and state-of-art of Support Vector Machine, and a detail introduction to the mathematical theory of Support Vector Machine, including deduction theory.2. Introduction and Summary of state-of-art and software implementation methods of online learning algorithm of Support Vector Machine, and hardware implementation algorithm of Support Vector Machine and the corresponding implementation thesis. And the summary of several hardware implementation methods.3. Since previous hardware architecture of SVM lacks of mentioning hardware implementation of kernel functions, which is the kernel of implementing online learning, two ways of hardware design of SVM kernel fuction are presented in this paper. These ways both supported multiple kinds of kernel functions and used serial operation style, which avoided the disadvantage of unflexibility to sample amount and excessive sensitivity of area to sample amount in previous structure. Moreover, these ways are compared in multiple perspectives, and discussed about the advantages and disadvantages. Experiment shows that one of them suits the situation of critical area limitation and supporting small amount of kernel functions, while the other one suits the situation of high-speed and supporting more amounts of kernel functions.