Learning rate adaptation achieved in unsupervised competitive learning: an application to noise cancelling

A fast unsupervised competitive learning rule is introduced for cancelling additive noise in slow-varying signals without using a priori knowledge about the underlying signal and noise distributions. The learning rule, called the fast boundary adaptation rule (FBAR), performs “online” adaptation of a scalar quantizer by maximizing the latter´s information-theoretic (Shannon) entropy. The learning rate is adapted using two identical FBAR-based networks, one with a fixed and another with a variable learning rate. The first is used as a reference against which the second is adapted. The performance is shown for stationary as well as non-stationary noise distributions added to speech and image signals