基于非负矩阵分解(NMF)的盲源分离算法采用乘性更新规则，但如何选择学习速率选择以及其对算法性能影响没有详细研究。对此，本文推导给出了选择不同学习速率时各种迭代更新公式，并对各种组合进行了大量计算机仿真实验，通过比较分析发现，有效的迭代更新公式的分母必须包含误差函数信息，分子分母的项数应尽可能平衡。 The iterative multipliable update formulas are used in blind source separation algorithms based on non-negative matrix factorization (NMF). However, the methods to select the learning rates and affect algorithms’ performance remain to be researched. This paper gives a derivation of different learning rates when selecting various iterative update formulas. A lot of computer simulations about these combinations are carried, and they show that a denominator of the effective iterative update formulas must contain information of the error function. In addition, its terms of denomi-nator and numerator should be balanced.
References
[1]
Lee, D.D. and Seung, H.S. (1999) Learning the parts of objects by non-negative matrix factorization. Nature, 401, 788-791.
Zdunek, R. and Cichocki, A. (2007) Nonnegative matrix factorization with constrained second-order optimization. Signal Processing, 87, 1904-1916. http://dx.doi.org/10.1016/j.sigpro.2007.01.024
Wang, S., et al. (2014) A K-L divergence constrained sparse NMF for hyperspectral unmixing signal. Proceedings of 2014 IEEE International Conference on Security, Pattern Analysis, and Cybernetics, Wuhan, 18-19 October 2014, 223-228. http://dx.doi.org/10.1109/SPAC.2014.6982689
