Nonlinear Blind Source Separation which is an extension of its more popular linear counterpart has gained increasing attention over recent years. Its development presents a more realistic approach due to the nonlinear mixing introduced by transmitter and receiver elements such as loudspeaker, amplifier and microphones. Though more accurate than linear models, it is also more complex and suffers from convergence issues. This paper proposes a polynomial neural network for blind nonlinear signal separation and also addresses the fundamental difficulty of non-unique solutions and slow convergence. Efficiency of the objective function is enhanced by a polynomial based model which is a flexible and more accurate fit. Coupled with reduced indeterminacy using additional constraints and improved convergence speed via adaptive learning rates, the proposed algorithm produces very promising results. Issues of convergence speed, accuracy and robustness against noise are investigated and results demonstrate the efficacy of the algorithm with adaptive learning rates.
|Title of host publication||Proceedings - 8th International Symposium on Signal Processing and its Applications, ISSPA 2005|
|Number of pages||4|
|Publication status||Published - 30 Jan 2006|
|Event||8th International Symposium on Signal Processing and its Applications, ISSPA 2005 - Sydney, Australia|
Duration: 28 Aug 2005 → 31 Aug 2005
|Conference||8th International Symposium on Signal Processing and its Applications, ISSPA 2005|
|Period||28/08/05 → 31/08/05|