Research On Auto-encoder Based On Radial Basis Function

In recent years,deep learning technology has made great progress in computer vision,natural language processing and reinforcement learning.While improving the performance of the algorithm,deep learning has not made corresponding progress in theoretical support.So far,deep learning is still regarded as a "black box".Due to the lack of theoretical guidance,it is becoming more and more difficult to make further improvement on the basis of existing theories.In this thesis,from the point of view of feature engineering,the existing auto-encoder model in deep learning is improved,and a radial basis auto-encoder model is proposed,and the Gauss function layer is introduced into the auto-encoder model.The model has certain theoretical advantages.First,the Gauss function has natural fuzzy logic attributes,which can be used as a bridge between the theory of fuzzy logic and the deep learning model.Second,Gauss function is a radial basis function,which is inextricably linked with the kernel method in machine learning.Thirdly,as an activation function,the Gauss function makes the network model have a certain ability of local approximation.The contributions and innovations of this paper are as follows:(1)Based on the kernel method,the classical auto-encoder algorithm is improved,the model of RBF auto-encoder is proposed,and the corresponding model training algorithm is given.(2)The Gauss function layer introduced by the new model can be transformed into the fuzzy rule layer.The introduction of the Gauss function layer is equivalent to the TS fuzzy reasoning system based on the fuzzy logic in the network.The experimental results also show that this combination is very effective.Finally,the proposed model is tested on three kinds of tasks(image reconstruction,high-dimensional data visualization,high-dimensional data reduction classification).Under the same experimental configuration,the new model outperforms the classical auto-encoder.Due to the particularity of the auto-encoder model,the evaluation index is diversified.Under the experimental settings in this thesis,the optimal reconstruction error of the new model on the more complex fashion-MNIST dataset is 0.022,which is lower than that of the classical auto-encoder(0.028).In the set classification experiments,the accuracy of the new model is 90.22%,which is higher than that of the classical self-encoder model(81.85%).From the effect of image reconstruction,the performance of the proposed algorithm is much better than that of the classical auto-encoder.The new model can reconstruct the details of the image,which is difficult to achieve by the traditional model.