中文摘要：

针对传统的自编码网络及其变体均采用均方误差作为重构函数对噪声不足,提出一种基于最大相关熵的堆栈稀疏自编码网络。该方法采用最大相关熵作为网络的重构函数,并且采用多层非线性映射层构建了一个多层网络,同时引入稀疏约束项。YaleB和AR人脸库实验结果表明,在训练样本有无噪声的情况下,该方法相比传统的自编码网络均具有更强的鲁棒性,且识别性能有所提高,学习到的特征更具表达能力。

英文摘要：

To overcome the problem of noise in the auto encoder network and its variants where mean square error is regard as reconstruction function, a stacked sparse auto encoder network is proposed, the maximum relative entropy is used as the re- construction function of network and a multi layer network with sparse constraint is constructed in the method. Experimen- tal results demonstrate that the proposed method is more and AR databases whether the training samples are noisy ance and the learned features are more powerful. robustness than the traditional auto encoder network on the YaleB or not noisy. In addition, it achieves better recognition perform ance and the learned features are more powerful.