Real-time on-line-learning support vector machine based on a fully-parallel analog VLSI processor

Abstract

An analog VLSI implementation of on-line learning Support Vector Machine (SVM) has been developed for the classification of high-dimensional pattern vectors. A fully-parallel self-learning circuitry employing analog high-dimensional Gaussian-generation circuits was used as an SVM processor. This SVM processor achieves a high learning speed (one clock cycle at 10 MHz) within compact chip area. Based on this SVM processor, an on-line learning system has been developed with the consideration of limited hardware resource. According to circuit simulation results, the image patterns from an actual database were all classified into correct classes by the proposed system. The ineffective samples are successfully identified in real-time and updated by on-line learning patterns. © 2012 Springer-Verlag Berlin Heidelberg.