Buffer Management in Online Kernel Machines

Machine learning has become an important tool for data scientists and engineers in recent years. Software and computer hardware can be used to perform pattern recognition tasks in many areas of research and have many machine learning applications. Online pattern recognition is an active area of research as the Internet of Things (IoT) has dramatically increased the volume of data requiring pattern recognition. Online machine learning has two advantages: Firstly, it can process data in a streaming fashion allowing for the transmission and storage of less costly meta-data versus raw data. Secondly, it can process data in edge-computing environments where resources are restricted in terms of processing capacity, data storage and power. It is the focus of this paper to investigate online training and pattern recognition by comparing the performance of three kernel algorithms. Online environments have limited memory and kernel machines require a buffer of captured data in order to form their decision function and learn from observed data. This paper proposes a novel algorithm for the replacement of old data within a kernel machine buffer. The experimental results of this paper show that a buffer replacement strategy that prioritizes replacement of examples in error has a positive impact on pattern recognition performance, and also a positive impact on total empirical loss exhibited by the kernel machine.