Hierarchical Few-shot Learning Based On Feature Correlations

The key issue faced by few-shot learning is how to mine effective information from limited data and establish a classification model with generalization ability,which can more accurately identify the categories of new task samples.The existing few-shot learning methods have made promising progress,mainly focusing on mining the correlations and information contained in data features to improve the model performance.Most of them assume that classes are independent of each other,ignoring the correlations between them.However,in practical applications,classes often do not exist independently but are related to each other,such as the multi-granularity correlation in the class hierarchical structure.The class hierarchical structure mainly adopts granular computing thinking,which originates from the human "from coarse to fine" cognitive mechanism.Therefore,it is an important challenge for few-shot learning how to use the multi-granularity correlations in the class hierarchical structure to guide the learning of feature correlations,simulating the idea of granular computing.In this dissertation,we focus on the research of hierarchical few-shot learning methods based on feature correlations,which mainly includes three aspects:1)Hierarchical few-shot learning method based on feature measurement via relation network:Most few-shot learning method assumes that classes are independent and ignores the correlations between them.To this end,we establish a hierarchical few-shot learning model based on feature measurement via relation network,which leverages the class hierarchical structure to guide the learning of feature similarity and class correlations.2)Hierarchical few-shot learning method based on global and local feature cross measurement:To tackle the problem that few-shot learning models are vulnerable to interference of background knowledge and suffer from information error,we propose a hierarchical few-shot learning model based on global and local feature measurement.From the feature and class perspectives,we take advantage of the global and local features and the class correlations to obtain more distinctive feature representations,which aims to eliminate the interference of background knowledge.3)Hierarchical few-shot learning method based on semantic guided feature contrastive learning:From the feature intra-class variance and inter-class commonness perspectives,we develop a hierarchical few-shot learning model based on semantic guided feature contrastive learning.It aims to solve the problem that it is difficult to distinguish the samples with large intra-class variance or large inter-class commonality.On the one hand,a semantic-guided supervised contrastive learning strategy is introduced to study the intra-class distinctive feature representation.On the other hand,using the granular computing idea of "from coarse to fine",we use a data hierarchical reasoning strategy to spread the commonness of inter-class features from coarse to fine to improve the model classification ability.To sum up,from the perspective of feature relationship,we leverage the granular computing idea to fully explore the variance and commonness among features to guide the fewshot learning from feature measurement to contrastive learning.Unlike traditional flat few-shot learning methods,this study constructs a hierarchical few-shot learning framework,which provides new ideas for the researches of few-shot learning in the future.