Kernel Size Adjustment Based on Error Variance for Correntropy Learning Algorithms

초록

The MCC (maximization of correntropy criterion) as one of ITL (information theoretic learning) criteria, has the inverse of kernel size squared in its slope for maximization, and this term causes the system instability. Recently, studies for adaptively adjusting the kernel size of the slope with the inverse of kernel size squared being removed. In those studies, however, the process of minimum error sample extraction employed for impulsive noise robustness leads the kernel size to zero after convergence, so that the weight adjustment cannot continue. In this paper it is proposed that without the minimum error sample extraction, an averaging and smoothing process on the absolute values of a block of error samples can create an appropriate kernel size. In the experiment, the proposed algorithm continues its weight adjustment even after convergence and yields enhanced learning performance by about 2 dB of steady state MSE with faster convergence speed compared to the conventional algorithm. © 2021, Korean Institute of Communications and Information Sciences. All rights reserved.

키워드