A novel unsupervised competitive learning rule with learning rate adaptation for noise cancelling and signal separation

A new ANN-based approach to adaptive noise cancelling and separating slow-varying signals is introduced. The network's weights are continuously modified using a fast unsupervised competitive learning rule, called fast boundary adaptation rule (FBAR), performing adaptive scalar quantization of the input signal. The rule maximizes information-theoretic entropy and yields a non-parametric model of the input probability density function. Contrary to classic unsupervised competitive learning, the author's system adapts its own learning rate, and hence does not require a 'cooling scheme'. Furthermore, contrary to most of the other noise cancelling approaches, the author's system does not require a priori knowledge or an explicit model of the joint noise and signal characteristics.<
>