A self-learning multiple-class classifier using multi-dimensional quasi-Gaussian analog circuits
Zhuoli Sun, Kyunghee Kang, Tadashi Shibata · 2010
A hardware-implementation-friendly classifier architecture having self-learning function has been developed for multiple-class classification. The similarity between two vectors is evaluated using a quasi Gaussian function which has been implemented by the summation of output currents from simple bump circuits. Binary weights are assigned to sample vectors and their values are determined by iteration similar to the SVM learning but in much simpler a way. Only one classifier is sufficient for N-class classification in contrast to N(N-1)/2 classifiers necessary in the SVM. The performances of the algorithm and circuits have been verified by software and SPICE simulations.