Research On Zero-shot Learning With Label Distribution Learning

As a new challenge in machine learning,zero-shot learning deals with the problem of using auxiliary information to predict unseen labels without corresponding training samples.Zero-shot learning can be divided into traditional zero-shot learning and generalized zero-shot learning according to the classes of samples in the test dataset,and can also be divided into three types according to the data used:class-inductive instance-inductive,class-transductive instance-inductive and class-transductive instance-transductive.In the previous study of zero-shot learning,the problem of class imbalance which is easy to occur with zero-shot is seldom taken into consideration.Moreover,the using of class auxiliary information is rough.The class auxiliary information is only used as the projection from attributes space to classes space.In the previous study of class-transductive instance-inductive zero-shot learning,it is difficult to solve the problem of bias towards seen classes by using the method of generating unreal samples.The main contributions of this paper are as follows:1.To solve the problem of class imbalance and insufficient use of auxiliary information in zero-shot learning,label distribution learning is introduced to the field of zero-shot learning.The embedding projection method based on the label distribution learning is proposed.With the help of class auxiliary information,the label distribution is calculated.The proposed solution alleviates the problem of class imbalance and achieves good results on several zero-shot datasets under the experimental setup of class-inductive instance-inductive zero-shot learning.2.To solve the problem of bias towards seen classes,a class knowledge transfering network based on label distributed learning is proposed.By dividing the zero-shot learning task into two subtasks,the problem of bias towards seen classes is alleviated.The proposed algorithm is tested on several zero-shot datasets and performs good.This paper consists into six chapters.Chapter 1 mainly introduces the concept,research status and research content of zero-shot learning and label distribution learning.Chapter 2 introduces the label distribution learning problem and the existing algorithm and evaluation method in detail.Chapter 3 introduces the zero-shot learning problem and the existing algorithm and evaluation method in detail.Chapter 4 introduces the embedding projection method based on label distribution learning.Chapter 5 introduces the knowledge transfering network based on label distribution learning.Chapter 6 summarizes the whole paper.