关键词: 深度学习,表达转换,数据降维,单层自编码器

Abstract: Deep learning frameworks and unsupervised learning methods have become increasingly popular and attracted the attention of many researchers in machine learning and artificial intelligence fields.This paper started from the “building-blocks” of deep learning methods and focused on the representation ability research of auto-encoders,especially on auto-encoders’ ability to reduce dimensionality and the stability of their representation ability.We expected that starting from the basis of deep learning can help us understand it better.Firstly,auto-encoder and restricted Boltzmann machine are two "building-blocks" of deep learning methods,both of which can be used to transform representation and be seen as relatively new nonlinear dimensionality reduction methods.Secondly,we investigated whether auto-encoder is a good representation transformation method under the context of understanding visual features,including eva-luating single-layer auto-encoder’s representation ability compared with classic methodology principal component analysis.Experiments based on original pixels and local descriptors demonstrate auto-encoder’s ability to reduce dimensiona-lity,the stability of its representation ability and the effectiveness of proposed AE-based transformation strategy.Fina-lly,future research direction was discussed.

Key words: Deep learning,Representation transformation,Dimensionality reduction,Single-layer auto-encoder