Decompositional independent component analysis using multi-objective optimization
Current approaches for blind source separation, such as independent component analysis (ICA), implicitly assume that the number of collected signals equals the number of sources. This assumption does not hold true in many real-world applications as in the case of electroencephalographic (EEG) data collected from the surface of a human's scalp, where independent EEG information is mixed with independent artifacts. This situation is abstracted in this paper by introducing the singers' party problem, where the number of signals collected from the party equals the number of singers. However, there are also a number of instruments playing at the party representing independent sources that need to be removed correctly to extract the voices of the singers. In this paper, we introduce a decompositional approach to project the sources found in ICA into a higher-dimensional space; providing the ability to separate local (singers) information from shared/global (instruments) information. The decomposition will also associate each component with a mixed signal, creating a bijective relationship between the mixed signals and the sources. The problem is formulated as a multi-objective optimization problem. We compare the pros and cons of two different multi-objective formulations of the problem and demonstrate that one of the formulations can effectively solve the singers party problem.