面向小样本约束的域适应分类算法

Domain Adaptation Algorithm for Few-Shot Classification Task

近年来,人工智能的相关应用被越来越细化到不同的应用场景,而对不同的应用场景都进行相应的数据收集,模型训练,模型调优等步骤需要消耗大量的时间精力会严重影响人工智能技术应用的效率.因此如何基于现有的成熟的训练过的模型迁移到其他应用场景是当前应用人工智能技术的关键问题.域适应算法主要研究将源域模型有效地迁移到目标域,这为上述问题提供了一个重要的解决思路.本文提出小样本对抗判别域适应算法,相对于无监督域适应算法能够在更严格的约束下-仅需要少量的目标域样本,在标准数据集上取得了优于对抗判别域适应算法(Adversarial Discriminative Domain Adaptation,ADDA)算法的表现,在单任务中最高提升幅度达16.9%.本文中,首先,提出了两种新的数据增强方法,以构建符合双域联合分布的图像以丰富样本多样性并填充特征空间,解决小样本约束下模型易过拟合到少量目标域样本的问题.接着,结合双域样本配对机制和ADDA算法,将以大量目标域样本为条件的无监督域适应算法改进为面向小样本约束的有监督域适应算法.在域适应过程中,引入类标签平滑损失来抑制过拟合现象,并结合度量学习中的最大平均差异度量,提出了新的域适应损失函数.同时,还提出了一种新的域分类判别器网络结构.最后,在对抗判别域适应算法的基础上增加了一个强化阶段,基于混淆矩阵对模型的分类性能进行强化提升.在困难数据集上的实验结果表明,仅使用少于5-shot的目标域样本,经提出的算法域适应训练得到的模型提升了26.3%~37.2%的分类准确率.

In recent years,the related applications of artificial intelligence have been more and more refined into different application scenarios,and the corresponding data collection,model training,model tuning and other steps for different application scenarios need to consume a lot of time and energy,which will seriously affect The efficiency of artificial intelligence technology application.Therefore,how to transfer to other application scenarios based on the existing mature trained models is a key issue in the current application of artificial intelligence technology.The domain adaptation algorithm mainly studies the effective transfer of the source domain model to the target domain,which provides an important solution to the above problems.This paper proposes a few-shot adversarial discriminative domain adaptation algorithm,which can achieve better performance than the adversarial discriminative domain adaptation(ADDA)algorithm on the standard datasets with different degrees of difficulty.The highest improvement rate is 16.9%in a single task.This does not only prove the effectiveness of the series of components proposed in this paper but also shows that the method proposed in this paper can bring a powerful impetus to the field of domain adaptation.In this paper,first,the two new data enhancement methods are proposed to enrich the diversity of samples and fill the feature space.The former is enhanced based on a mixup in the global perspective,and the latter is enhanced based on a cutmix in the local perspective to construct the images which conform to the joint distribution of the source domain and the target domain.This is to solve the problem that the model is easy to overfit to a small number of target domain samples under the constraint of few-shot.Then,combining the dual-domain sample pairing mechanism and the ADDA algorithm,the unsupervised domain adaptation algorithm is improved to a supervised domain adaptation algorithm oriented to a few-shot constraint.Compared with unsupervised domain adaptation methods that require a large amount of unlabeled target domain data,the few-shot condition requires only a very small amount of labeled target domain data,which greatly relaxes the constraints on data collection and makes the method proposed in this paper have a broader prospect.Furthermore,in the domain adaptation process,the concept of class label smoothing is introduced to modify the naive loss function to suppress the overfitting problem,and combined with the maximum mean discrepancy metric in metric learning,a new domain adaptation loss function is proposed to increase the separation degree of features.At the same time,a new domain classification discriminator network structure is proposed to stabilize and speed up the model training.Finally,on the basis of the adversarial discriminative domain adaptation algorithm,an enhancement stage is added to enhance the classification performance of the model based on the confusion matrix.Experimental results on difficult datasets show that using only 5-shot(or less)target domain samples,the model after the training of the proposed domain adaptation algorithm improves the classification accuracy by 26.3%-37.2%,which proves the powerful performance of the method proposed in this paper.