Semi-Supervised Learning Based On Manifold Learning And Its Application

Semi-supervised learning has always been an important and challenging research topic in the field of machine learning and computer vision.Based on the given dataset,traditional methods aims to learn the mapping function from the feature space to the target space,which always followed by a classification or regression task.However,quantities of labeled examples are impractical in the real-world problems.Semi-supervised learning focuses on how to effectively utilize the few labeled data and massive unlabeled data during training stage,and expect the performance of decision function can be comparable with the decision function that trained with full-labeled data.The existing incomplete labeled data distribution can be roughly divided into two types:(i)each class consists of labeled and unlabeled data,we call this problem as classical Semi-Supervised Learning(SSL).(ii)Only positive(or negative)and unlabeled examples exists during training,such problem is named as the PU Learning.In this paper,we first introduce the background and the development of semi supervised learning.Then,we show the classical methods in this research field.Afterward,we propose two robust algorithms,and extensive experiments on real-world datasets demonstrate that our methods is able to achieve the state-of-the-art performance.Finally,for the actual scenario,this paper introduce a brand new semi-supervised framework,which can be applied to signal propagation with deep learning.The innovations of our work are summarized as follows:1.Traditional semi-supervised learning focuses on how to utilize the available unlabeled data to improve the training quality.However,their estimations of the distribution revealed by unlabeled examples might not be accurate as the classes information of unlabeled examples are unknown.Meanwhile,previous work seldom consider to introduce additional auxiliary information to improve the performance.Inspired by the framework of Learning Using Privileged Information(LUPI),we propose to introduce an intelligent “teacher”which can utilize the privileged information(i.e.the explanations of training set)during training stage to improve the performance of SSL.Our method can significantly improve the performance of classifier with only a small amount of label cost.2.Most existing models follow the scheme of risk minimization and adopt the shallow learning framework,resulting in highly limited learning capability especially on large-scale datasets.To address the above issues,we propose a new method dubbed Deep Positive and Unlabeled learning with Manifold Regularization(DeepPUMR),which builds a deep neural network and also introduces the manifold regularization for PU learning.Specifically,we introduce a new risk estimator named Manifold Nonnegative Risk Estimator(MNRE),which is able to capture the underlying manifold properties hidden in the entire dataset,and embed it into the high-performing convolutional neural network(CNN)ResNet.Substantial experimental results demonstrate that the proposed Deep-PUMR obtains superior performance to the state-of-the-art PU learning methods on a variety of large-scale datasets.3.Existing semi supervised learning is only adapted to tackle the image or text information.Combining with the real-world application,this paper design a new deep learning framework called Semi-Signal Network(SSig-Net).The merits of SSig-Net are two-folds: firstly,we use both labeled and unlabeled data during training,which immensely improves the accuracy across different scenes.Secondly,to make the inference more suited for the signal propagation,we redefine the convolution operator from shifting to radiate.Our method greatly improve the prediction accuracy on the real-world dataset.