Unlabeled Data Aided Deep Learning Techniques Researches

Deep learning techniques have been applied to many areas.Training deep models usually requires a large amount of labeled data,yet it consumes human and material resources to get the label information in many real applications.Therefore,how to utilize unlabeled data to improve the performance of deep learning is an urgent problem.This thesis focuses on it and achieves the following innovations:1.Conduct a study on utilizing unlabeled data to help improve the performance for supervised learning of deep learning techniques,and propose the method TDNN(Tri-training Deep Neural Network),which incor-porates the semi-supervised learning mechanism Tri-training into a deep model.The experimental results show that the method in this thesis can improve the performance for supervised learning of deep learning.2.Conduct a study on utilizing unlabeled data to help improve the performance for transfer learning of deep learning techniques,and pro-pose the method SDA-TCL(Semantic Domain Alignment and Target Classifier Learning),which jointly optimizes semantic domain alignment between two do-mains and the loss function in the target domain.The experimental results show that the method in this thesis can improve the performance for transfer learning of deep learning.3.Conduct a study on utilizing unlabeled data to help improve the performance for metric learning of deep learning techniques,and propose the method SDVN(Stable Deep Verification Network),which jointly optimizes the stability of output similarity with noise and the loss function for supervised information.The experimental results show that the method in this thesis can improve the performance for transfer learning of deep learning.