Local And Non-local Regularization For Semi-Supervised Deep Learning

With the successful application of deep learning inmany fields, the regularization for deep neural network has also done a lot of research by embedding a regularizer to the loss function, thus the feature is more beneficial to the subsequent application. However, those regularization methods have posed serious challenges to the efficiency of the classification task.In view of the above problems, we devise a new regularized method, the method integrate the local and non-local constraints in the labeled and unlabeled s amples to extract abstract features that are effective for preserving the class-separabilit y in the raw feature space. For the labeled data, we make use of the label to define the pairwise proximity matrix, then the topological regularizer can be obtained by minimi zing intraclass(local) compactness and maximizing interclass(non-local) separability.For the unlabeled data, we use the average distance from one sample to others as the t hreshold to select its neighbors apart from the non-local samples. Then the topological regularizer for unlabeled samples is to simultaneously maximize the nonlocal scatter and minimize the local scatter. By intergrating local/non-local topological regularizer for both labeled and unlabeled data, our discriminant regularization can extract feature s that are more suitable for classification purpose.In conclusion, we present a solution for local and non-local regularization for semi-supervised deep learning, a novel regularization algorithm is provided in this paper. The proposed method is investigated on several publicly available image datasets. The extensive experimental results demonstrate that the method is effective in feature extraction, leading to promising image recognition performance.