Research On Robustness And Scalability Of Label Distribution Learning

Label Distribution Learning(LDL)is a new supervised learning paradigm and has been applied to many applications successfully,for example,facial age estimation,facial expression recognition and movie rating estimation.However,the robustness and scalability of LDL have not been exploited in existed researches.The robustness of LDL means when training sets are unusual,LDL will have better performance by using label correlation.The unusual training sets include following cases.First is zero-shot learning,that is in some tasks,it's hard to obtain instances of certain classes,thus there are no training instances for some classes during the progress of training.Second,no matter the labels are annotated by human or acquisition devices,noisy labels will exist inevitably and these noise will influence model's performance.The scalability of LDL is solving large-scale machine learning problems based on existed LDL methods.Large-scale machine learning problems means large-scale training samples and large-scale label space.This paper focus on the second problem,that is it's hard to learn a good mapping from instance to label directly in large-scale ordinal label space.The target of this paper is to deal with above problems and study the robustness and scalability of LDL.The main contributions of this paper: 1.This paper applies LDL into ordinal zero-shot learning through generating supervision intensity distribution by using label correlation and expanding supervision information from seen labels to unseen labels,which proves good robustness of LDL;2.This paper uses LDL to process ordinal noisy labels and studies the method based on label distribution can alleviate the influence of annotated noisy labels through utilizing the adjacent labels around the ground truth label preferably,which further proves the robustness of LDL;3.This paper extends the LDL and proposes the hierarchical label distribution learning algorithm to deal with the problem which has large-scale ordinal label space.Through handling the large-scale ordinal label space,it demonstrates the scalability of LDL.This paper consists of six chapters.The first chapter introduces the research status of LDL and the research target of this paper.The second chapter introduces problem definition of LDL and existed algorithms and evaluation methods of LDL.The third chapter introduces the research on ordinal zero-shot learning by using LDL.The forth chapter introduces applying LDL to deal with ordinal noisy labels.The fifth chapter introduces hierarchical label distribution learning and verify the scalability of LDL.The last chapter is the summary and future outline of our work.