Optimizing Deep Learning Algorithm Based On Noisy Autoencoder

Deep learning is an emerging approach within the machine learning research community.Deep artificial neural networks research aims at discovering learning algorithms that discover multiple levels of distributed representations,with higher levels representing more abstract concepts.The performance of many machine learning methods is heavily dependent on the choice of data representation or feature on which they are applied.Autoencoder has been successfully used as an unsupervised learning framework to give useful representations in deep learning tasks.Sparse autoencoder,denoising autoencoder and contractive autoencoder are the optimized algorithms of traditional autoencoder based on the regularization techniques.Due to the drawbacks of conventional algorithms on the detection precision and robustness,this thesis presents a new deep neural network based on noisy autoencoder.Both of input layer and hidden layer of the traditional autoencoder are partial corrupted by random noise.During training,the goal is to reconstruct the input data and minimize the reconstruction error.With unsupervised greedy layer-wise pre-training strategy,the noisy autoencoder can be stacked to build a deep learning architecture,that avoids the local optima and diffusion of gradients.Moreover,the impact of different kinds of hidden activation noise on the classification performance is analyzed with experimental results.Furthermore,our deep learning method significantly improves learning accuracy when conducting classification experiments on benchmark data sets.