一种互信息梯度不变的非线性特征提取方法*

针对线性的互信息特征提取方法，通过研究互信息梯度在核空间中的线性不变性，提出一种快速、高效的非线性特征提取方法。该方法采用互信息二次熵快速算法及梯度上升的寻优策略，提取有判别能力的非线性高阶统计量；在计算时避免传统非线性特征提取中的特征值分解运算，有效降低计算量。通过UCT数据的投影和分类实验表明，该方法无论在投影空间的可分性上，还是在算法时间复杂度上，都明显优于传统算法。

Invariance of mutual information gradient based nonlinear feature extraction

This paper proposed a fast and effective method of nonlinear feature extraction by studying the linear invariance of mutual information gradient in the linear mutual information feature extraction. It employed a fast algorithm for mutual information and gradient ascent which avoid the eigenvalue decomposition of the traditional nonlinear transformation. In this way, the extracted features could reflect the characteristics of discriminative higher-order statistics, and effectively reduce the computational complexity. Experiments with the UCI read data show that the proposed approach performs well in projection and classification performance, and is better than traditional nonlinear algorithms for the time complexity.