Based on the minimum mean square error of principal component analysis and principal component neural network is an effective dimensionality reduction of multivariate statistical techniques, they extract the main element contains the system maximum variance approximation model non-Gaussian stochastic systems should contain maximum entropy system, but contains the maximum variance does not necessarily contain the maximum entropy. This paper presents a minimal residual entropy for the general index nonlinear principal component neural network model, and gives an approximate calculation method based on Parzen window density function estimation of entropy and network learning algorithm then analyzed from the perspective of information theory, the Gaussian random system based on minimum residual entropy and minimum mean square as an indicator of the primary element network learning outcomes consistent. Finally, simulation effectiveness of the method, and with comparative analysis based on minimum mean square error calculation principal component analysis and principal component neural network method.
|Translated title of the contribution||A principal component neural network with minimum error entropy for dimension reduction of non-Gaussian systems|
|Number of pages||4|
|Publication status||Published - 29 Dec 2005|
- PCA neural networks
- minimum residual entropy
- minimum mean squared error