Research On Few Shot Learning Based On Metric Learning

Few shot learning is a task of learning from a few examples,which poses a great challenge for current machine learning algorithms.One of the most effective approaches for few shot learning is metric learning.In particular,the goal of metric-based approaches is to learn a mapping function.The quality of learned mapping determines the effectiveness of metric learning algorithms.However,due to the scarcity of samples,it is extremely difficult to learn a high-quality mapping function.Based on that,this paper proposes few shot learning with margin,which improves the quality of the learned embedded representation.To introduce the margin,this paper proposes multi-way contrastive loss based on the characteristics of few shot scenes.So the few shot learning model can learn a more discriminative metric space.And generalization errors can be reduced.Few shot learning with margin is a common framework that can be combined with various metric-based few shot learning models.In this paper,few shot learning with margin is integrated into two existing models,namely prototypical network and matching network.In addition,existing algorithms do not consider the characteristics of the data distribution when classifying,which hinders the effectiveness of the algorithm.This paper takes advantage of a graph regularization-based relation propagation framework that infers the relationship between unknown examples by combining the manifold of the sample distribution with the known relationship among the examples.The objective function of the framework is a convex optimization problem,so the global optimal solution can be obtained.Based on this framework,this paper proposes a graph regularization-based few shot learning algorithm,which integrates graph regularization into the task of few shot learning.Due to the existence of graph regularization,the algorithm can fully consider the characteristics of data distribution,thereby improves the accuracy of classification.In this paper,a series of experiments on standard datasets demonstrate the effectiveness of few shot learning with margin and graph regularization-based few shot learning.The experimental results show that the few shot learning with margin is obviously better than the original model,and graph regularization helps on the classification stage of few shot learning.