Unsupervised separation of discrete sources with a combined extended anti-hebbian adaptation

In the classical methods of unsupervised source separation, the a priori hypothesis is independence of sources. In certain applications, there is some additional knowledge on the sources (statistics, distributions, alphabet…). It is the case with discrete sources with known alphabet. Then we can improve separation. Initialization of adaptation is done according to some known algorithm, e.g. thanks to an extended anti-Hebbian algorithm, provided there are not less sensors than sources. As soon as the separation performance index has reached some preassigned level, a second part which involves the output decision error is introduced in the increment. In a noiseless environment, this method allows complete cancellation of steady state adaptation fluctuations and perfect source recovery.