Research On Active Learning Algorithms In Continual Learning Framework

With the rapid development of computer technology,machine learning has achieved outstanding performance in many fields.New machine learning algorithms are developed continuously,but its success heavily relies on the large number of labeled training samples.However,for many professional fields,data annotation is not only tedious and time consuming,but also demanding specialty-oriented knowledge and skills,which are not easily accessible.To significantly reduce the cost of annotation,we propose a novel continual learning framework called CLBSS.CLBSS is based on active learning method and the main purpose of active learning is to reduce the number of labeled samples by selecting "the most valuable" samples from a large number of unlabeled samples through active learning sample selection algorithm.CLBSS uses active learning algorithm to select the best subset from the unlabeled samples for manual labeling and then completes model updating continuously.CLBSS mainly consists of three modules:base classifier,active learning sample selection algorithm and labeled sample sampling algorithm.Base classifier is the main body of CLBSS.Active learning algorithm and labeled sample selection algorithm both depend on base classifier.Different base classifiers are usually used according to different classification tasks.In this paper,the task scenario is set as audio classification.Audio classification usually first converts the original audio features into spectrum maps,and then do classification on them.Thus,the convolutional neural network(CNN),which designed for image classification,is the first choice of the base classifier.Active learning sample selection algorithm is the core part of active learning process.Common active learning sample selection strategies are mainly based on the "most discriminative" or the"most representative" criteria.The "most discriminative" criteria considers characteristics of the current model,selecting the most "fuzzy" samples for manual labeling.The "most representative"criteria takes the characteristics of data structure into account.CLBSS adopts a new active learning sample selection strategy which combines the "most discriminative" and "most representative"criteria,which draws on the advantages of the two strategies,making the selected samples more reasonable and more conducive to the improvement of classifier ' s performance.The labeled sample selection algorithm downsampled the samples learned by the classifier and achieved a good balance between learning performance and efficiency.With the increase of iteration rounds,the number of labeled samples will also increase.Downsampling of labeled samples can alleviate the problem of data expansion.However,if we do not use the labeled data that we have learned at all and only use the new labeled data in the current batch,it will cause catastrophic forgetting problem,that is,the classification performance of the data that the classifier has learned will decline sharply.Thus,using some of the data we've learned can also help alleviate catastrophic forgetting.The experiment results on two different audio datasets demonstrates that the active learning algorithm based on syncretic strategy can improve the performance of the classifier more robustly and rapidly.Comparing with other continual learning frameworks,CLBSS can avoid unnecessary computational costs and reduce storage requirements by downsampling labeled data.