Adaptive blind sources separation based on unsupervised learning

The separation of the independent sources from an array of sensors without the knowledge of channel characteristics is a fundamental problem encountered in many applications. For example, the signals received by antenna sensors in a communication system is often a deterministic linear transformation of statistically independent signals, {dollar}ssb{lcub}i{rcub},{dollar} called sources. Processing is then needed to restore these original source signals, without the knowledge of the linear transformation.; When the channel characteristics or the statistics of sources are slowly changing, adaptive blind sources separation is necessary. It is well known (28) that the source signals can be recovered by achieving pairwise independence. However, the realization of pairwise independence requires a knowledge of the joint probability distribution, which can not be obtained in practical applications. This dissertation, first, presents a new theorem which requires only 2nd-order and 3rd-order pairwise statistics to recover the source signals. Second, neural networks with local adaptive learning rules are constructed to implement the proposed theorem. Third, two-stage neural network, which is an on-line implementation of FOBI algorithm proposed by Cardoso, is presented.