On the convergence of “A self-supervised vowel recognition system”

摘要

A self-supervised learning algorithm based on the concept of guard zones was developed by Pal et al.(1) for studying the adaptive ability of a recognition system, starting with non-appropriate representative vectors. Guard zones were used to discard unreliable (doubtful) samples from the parameter-updating programme, so that the convergence does not get affected. The algorithm was implemented with success on speech data but no proof of convergence was provided.The present paper investigates the convergence of this algorithm, using some results on multidimensional stochastic approximation. It is shown that the estimates of the parameters converge strongly to their true values under certain conditions provided the guard zones are effective in discarding mislabelled training samples.