Visual Data Classification Based On Active Learning And Semi-supervised Learning

With the evolution of socioeconomic status and information technology,there have been massive amounts of data in all walks of life.Processing and analyzing the data by machine learning method plays an important role.As we all know,number and quality of labeling samples are major factors affecting learning results.But,it always pays to obtain labeling samples.This leads to shortage of labeling samples in many machine learning tasks.Active learning,transfer learning and semi-supervised leaning are three kinds of approaches to solve this problem.There have many works in these fields,however,some issues still need to be studied.For example,When there are some labeling samples of related domain,the combining of transfer learning and active learning can be used to alleviate the lack of labeling instances.In this scenario,we need to consider how to combine active learning and transfer learning with their full advantages to handle the problems of insufficient labeling samples.When there are limited labeling instances with abundant unlabeled instances,graph-based semi-supervised can be used to eliminate the shortage of labeling samples.In such situations,we need to study how to overcome the sensitivity to labeling noises in semi-supervised learning.For these issues,we have proposed two kinds of innovative methods.Firstly.We combine active learning and transfer learning,and propose an active transfer learning method based on Shannon entropy.This method combines the advantages of active learning and transfer learning,and it not only reduces the cost of annotating instances,but also utilizes the related knowledge in source domain,resulting in accurate mod-els with a handful of labeling samples.Secondly,we propose a robust graph-based semi-supervised learning method via maximum correntropy criterion.This method in-troduces supervised information to the regularization item,and forces the supervised information to constraint the predicted values of unlabeled samples through the struc-ture of the graph,thus making the predicted values close to the ground truth.In order to reduce the negative impact of labeling noise,maximum correntropy criterion is in-troduced to measure the similarity between labels(predicted labels and ground truth),and the the influences of labeling noise to classification results are suppressed.We have conducted extensive experiments compared with some advanced related methods in numbers of visual datasets,the effectiveness of the proposed methods are proved through experimental analysis and theory demonstrating.