One of the greatest challenge of machine learning is the curse of dimensionality. Tolessen this burden on the standard algorithms, a set of dimensionality reduction methods havebeen developed. Among them, the neural network based dimensionality reduction algorithms,which includes the sparse auto-encoder and deep neural networks, have shows their greatability and potentiality against complex real-data problems.In this paper, we first survey on the neural network based dimensionality reductiontechniques. In this part, a down-top investigation on the basic principle is first performed.Then, we analysis the interconnection between neural network based dimensionality reductionalgorithm and other state-of-art techniques, such as sparse coding and deep learning.Furthermore, a comparison between neural network and other dimensionality reductionmethods is done.We also adapted the neural network dimensionality reduction algorithm into TWO areaas application:FIRSTLY, on the aspect of Approximate Nearest Neighbour Searching, we developed aneural network based 2-phase nearest neighbour searching strategy, which is based on theconqure-and-divide idea and making full use of the hierarchical features learned by a neuralnetwork. Our algorithm makes it possible to achieve low dimensionality and considerablequality of features, bringing double improvement on accuracy and speed.SECONDLY, on the aspect of Iris Recognition, we proposed a fast framework for irisrecognition, which consists of a rotation invariant feature extraction module by PCET and afollowed dimensional reduction module implemented by a deep neural network.The experiment on both algorithms show that our adapted neural network can greatlyreduce the dimensionality whereas retaining high performance. |