Design And Implementation Of Semi-Supervised Continuous Learning Framework

Nowadays,with the rapid development of machine learning,it has been applied to many fields and has achieved excellent performance.However,increasing actual requirements have put forward higher requirements for machine learning.Traditional machine learning algorithms rely on a large amount of manually annotated data to build models,but not only does it take a lot of time to annotate the data,but it also requires many human resources with professional knowledge.Therefore,it often happens that there is only a small amount of labeled data,and a large amount of data is unlabeled.In addition,the data in the actual scene often arrives continuously,and the definition of the data category will change with time,that is concept drift.In order to reduce the waste of data resources and further utilize unlabeled data,this paper explores the relationship between sample categories and features,and proposes a local-representation-coefficients-based nearest neighbor classification LRCBNN.LRCBNN uses all labeled and unlabeled data to find the nearest neighbors of the query samples by mining the overall distribution of the data feature space,and then uses sparse representation classification algorithms to obtain classification results.In order to alleviate the scarcity of labeled data and concept drift in real scenes,this paper takes the audio classification task as an example to design and implement a semi-supervised continuous learning framework SSCLF applied to the classification model,and introduces the implementation details of the framework software.SSCLF first uses a small amount of labeled data to initialize the classification model,and then every time a batch of new data arrives,after active learning and active learning enhancement,the existing model will be trained and updated to continuously improve performance.Among them,the active learning enhancement algorithm uses the neighbor search method of LRCBNN to filter the neighbor samples from the unlabeled data and pseudo-label based on the knowledge obtained by active learning.Then SSCLF adopts a semi-supervised training method which iteratively uses pseudo-labeled data and labeled data.This method can reduce the dependence on the amount of labeled data,effectively alleviate catastrophic forgetting.Finally,experiments on two audio datasets verify that SSCLF can improve model performance faster than other continuous learning frameworks without increasing the cost of manual annotation.