面向小样本学习的动态分布校正方法

Method of Dynamic Distribution Calibration for Few-shot Learning

近年来,机器学习在多领域取得巨大成功。在现实世界的真实场景中,由于采集环境复杂或标注难,使得可用于训练的标准样本很少,导致机器学习模型往往出现过学习或欠学习的问题。所以,小样本学习是具有挑战性的机器学习问题。近期人们提出分布校正方法,他们假设特征表示中每个维度都服从高斯分布,利用基类的特征分布来校正新类的特征分布。然而该方法易引入负迁移现象,并且易于淹没新类本身的特征分布。为此,该文提出一种动态分布校正方法解决分布校正过程中的负迁移问题。首先,基于阈值动态地选择近邻基类和远域基类;其次,新类样本的特征经过幂变换的操作后,增加特征标准化处理来消除不同量纲之间的差异;最后,该方法引入参数调节迁移分布与新类特征原分布之间的比例关系来实现新类的特征分布校准。通过在常规数据集miniImageNet和CUB上与最新算法和传统算法的大量对比实验表明,该方法可以有效提升小样本分类任务的性能。

In recent years,machine learning has achieved great success in many fields.In the real scene of the real world,due to the complex collection environment or difficult annotation,there are few standard samples available for training,resulting in the problem of over learning or under learning of machine learning models.Therefore,few-shot learning is a challenging machine learning problem.Recently,people have proposed a method of distribution calibration,which assume that each dimension in the feature representation obeys a Gaussian distribution,and use the feature distribution of the base classes to calibrate the feature distribution of the novel classes.However,this method is easy to introduce the phenomenon of negative migration,and it is easy to submerge the characteristic distribution of the novel classes itself.Therefore,we propose the dynamic distribution calibration to solve the problem of negative migration in the distribution correction method.Firstly,the base classes of the nearest neighbor and the base classes of the far domain based on the threshold value are selected dynamically.Secondly,the standardization processing is added to the sample characteristics of the novel classes after power transformation to eliminate the differences between different dimensions.Finally,the method introduces parameters to adjust the proportional relationship between the migration distribution and the original distribution of the features of the novel classes to achieve the calibration of the feature distribution of the novel classes.A large number of comparative experiments with the latest algorithm and the traditional algorithm on the conventional data sets miniImageNet and CUB show that the proposed method can effectively improve the performance of few-shot classification tasks.