Blind signal separation using fixed overcomplete basis function dictionaries

A solution for achieving blind separation for underdetermined systems is to use an overcomplete basis function set that has the ability to span all possible inputs. Ideally, such a basis would be learned for each set of inputs but this is computationally expensive. A less processor intensive system is shown using a fixed dictionary of basis functions learned from existing sources and reduced using a correlation-based method. The relation between dictionary size and separation performance for underdetermined scenarios is examined and we demonstrate that a reduced dictionary can produce comparable results using less computational power.