Hierarchical Learning With Few-Shot Based On Knowledge Granularity

With the rapid development of internet technology,the speed of data acquisition,collection and statistics is increasing.The number of samples and characteristic dimensions of the obtained data have increased explosively,and the types of data have changed from hundreds of thousands to hundreds of millions or even more.For instance,the MNIST dataset designed in 1998 has 60000 samples,and the samples are handwritten digits.The Image Net dataset designed in 2012 has more than 14 million samples,covering more than 20000 categories,where has more than one million pictures,clear category labels,and object position labels of images.However,these phenomenon have resulted in a series of problems of unbalanced data distribution:(1)The distribution of sample categories is uneven where the number of different categories varies greatly,and the number of different subcategories of the same category varies greatly;(2)The continuous emergence of new categories challenges the traditional classification methods,which greatly increases the cost of manual labeling of samples;(3)It is difficult to effectively model for the classes with scare samples,resulting in the difficulty of sample category classification.For the classification task of uneven data distribution,this thesis proposes a hierarchical few-shot learning method based on knowledge granularity by fully excavating the relationships between different knowledge granularities and category relationships between samples.The main contents include:(1)Few-Shot Learning via Relation Network Based on Coarse-Grained Granulation.Aiming at the emerging few-shot data of new categories,this thesis leverages the coarse granularity of categories for similarity matching and constructs a few-shot learning model of relational network based on coarse granularity granulation.First,the category knowledge structure with different granularity is constructed through coarse-grained granulation.Then,the relationship module is used for similarity matching,and the relation scores between new samples and different categories of coarse-grained are calculated.Finally,a few-shot learning method of relationship network based on coarse-grained granulation is proposed.(2)Hierarchical few-shot learning based on top-down mechanism with stop strategy.For the classification of sample data with different knowledge granularities,there are not all coarse granularities have correct classification values.The samples with different knowledge granularities are utilized to construct the knowledge graph for auxiliary classification.The problem of inter-level error propagation is avoided through the stop strategy.Then,the sample classification results are revised from top to bottom by using the top-down mechanism.Finally,a stop strategy hierarchical classification based few-shot learning method based on the top-down mechanism is proposed.