Zero-Shot Transfer Learning Based On Deep Models

Deep learning has achieved unprecedented development in the past decade,achieving great success in various fields such as computer vision.However,existing deep learning methods rely on a large amount of labeled data,which limits their applications in many real-world scenarios where labeled data is limited.Transfer learning is a popular learning paradigm to solve the problem of limited data in the target task.In transfer learning,deep models transfer the knowledge learned from data-rich source tasks to data-limited target tasks,thereby reducing its dependence on massive labeled data of the target task.This thesis focuses on a very important and common problem in transfer learning,called zeroshot transfer learning.Different from traditional transfer learning,zero-shot transfer learning aims to solve the problem of how to effectively transfer the knowledge learned from the source task to the target task without any label information or labeled data.Specifically,this thesis studies three types of zero-shot transfer learning,including transductive zero-shot transfer learning where only unlabeled data are available for the target task,inductive zero-shot transfer learning where no data are available for the target task,and model transferability measuring with no training samples.The details are as follows:1.In terms of transductive zero-shot transfer learning,trained models in existing methods usually suffer from the projection shift problem,where the trained models easily misclassify target data into the source categories.To alleviate the projection shift problem,this thesis proposes an unbiased embedding for transductive zero-shot transfer learning.To fully leverage the unlabeled target data,this thesis further proposes the Quasi-Fully-Supervised Learning to alleviate the projection shift issue and meanwhile to maximize the performance on both the source and the target tasks.2.In terms of inductive zero-shot transfer learning,the manually defined attribute space is usually under-complete to some degree,which hinders the transfer of knowledge from the source categories to the target categories,resulting in the low accuracy and high risk in practical problems.This thesis proposes an inductive zero-shot transfer learning method based on automatic attribute augmentation.This method exploits dictionary learning to automatically mine the remaining attributes other than artificial attributes to build a more complete attribute space.Additionally,this thesis proposes to solve the zero-shot recognition problems in the manner of selective classification,so as to reduce the prediction risk in practical problems.3.In terms of measuring the model transferability with no data,existing transfer learning methods cannot select the optimal pre-trained model when multiple pre-trained models are available,and cannot select the optimal layer for feature extraction as there are many layers in a deep model.This thesis proposes an explainable and comparable representation of deep knowledge,dubbed as DEeP Attribution gRAph(DEPARA),to investigate the transferability of knowledge learned from heterogeneous tasks.Equipped with DEPARA,the model transferability and the layer transferability can be directly measured,even if no training data are given.This thesis organizes the research on the aforementioned zero-shot learning problems in the order of the amount of data from large to small,and the tasks from easy to difficult.The promising experimental results indicate that this study is a good complement and improvement to the field of zero-shot learning.