Hierarchical Few-shot Learning Based On Semantic Knowledge Transfer

With the development of deep learning,models based on massive data have achieved excellent results in classification issues.But it is difficult to classify for few-shot learning problems with a small number of samples.Humans can apply knowledge learned in the past to the learning of new classes,enabling them to acquire the ability to identify new classes through a few samples.This method of knowledge transfer brings new ideas to few-shot learning,but there are still some limitations:(1)The traditional flat few-shot learning method ignores the semantic structure knowledge between classes;(2)There is not only semantic hierarchical structure knowledge between data,but also more knowledge such as data features to be mined.This causes the previous few-shot learning methods to perform poorly on the classification tasks with a few samples.In this thesis,we fully excavate the relationship among the data to obtain abundant knowledge for the few-shot learning task.We design the hierarchical few-shot learning based on semantic knowledge transfer.The main contents are as follows:(1)Hierarchical few-shot learning model based on semantic knowledge transfer.Traditional few-shot learning models only utilize the flat data information and ignore the existing hierarchical knowledge structure among classes.We take advantage of a tree-structured knowledge graph to facilitate the classification results.We consider a tree-structured class hierarchy according to the semantic information among classes as a knowledge graph to alleviate the lowdata problem.Finally,a hierarchical few-shot learning model based on semantic knowledge transfer is proposed.(2)Hierarchical few-shot learning with multi-perspective knowledge transfer.There is not only semantic hierarchical structure knowledge among data,but also more knowledge such as feature knowledge to be mined.Firstly,the attention mechanism is used to obtain highly focused regions of sample features,which play a much greater role in classification than other regions.Secondly,the highly focused areas among samples are migrated to each other to obtain more samples for training.Then,we weigh the participation of semantic knowledge and feature knowledge in loss function.Finally,a hierarchical few-shot learning with multi-perspective knowledge transfer is constructed.