基于自编码神经网络的手写体数字识别中关于特征表达的研究

研究多种改良的自编码神经网络（Autoencoder），如稀疏（Sparse）、噪声（Denoising）、权值对称（Tied Weight）。探究这些自编码神经网络的改良在图像特征表达中的原理。将方法应用到手写体数字的识别中，通过设置各种改良自编码神经网络的参数取值并且对比各种改良自编码神经网络的特征表达效果，证明改良自编神经网络的理论原理。实验证明稀疏和噪声对于自编码神经网络性能具有较大提升。

Study on Feature Representation in Handwritten Numeral Recognition Based on Autoencoders

This paper studies on several advanced Autoencoders like Sparse Autoencoder, Denoising Autoencoder, and Tiedweightstrategy for Autoencoder; explores on the principal of improving the feature representation with these Autoencoders, whichhave been certificated in the handwritten numeral recognition. The performances of feature representation with these Autoencodersare compared by setting the Autoencoders parameters. Experimental results show that the Sparse and Denoising strategies have greatimprovment to the performance of Autoencoders.