Maximum Independent Component Analysis with Application to Non-linear Temporal Signals

Chunming Zhang (University of Wisconsin-Madison)

In many scientific disciplines, finding hidden influential factors behind observational data is essential but challenging. The majority of existing approaches rely on linear transformation, i.e., true signals are linear combinations of hidden components. Motivated from analyzing non-linear temporal signals in neuroscience, genetics, and finance, this paper proposes the Òmaximum independent component analysis" (MaxICA), based on max-linear combinations of components. In contrast to existing methods, MaxICA benefits from focusing on significant major components while filtering out ignorable components. A major tool for parameter learning of MaxICA is an augmented genetic algorithm. Extensive empirical evaluations demonstrate the effectiveness of MaxICA in either extracting max-linearly combined essential sources in many applications or supplying a better approximation for non-linearly combined source signals, such as EEG recordings analyzed in this paper.


Back to Day 1