Research On Cross-Domain Meta-Learning

Currently,meta-learning methods have been able to tackle the few-shot problem well,but when the few-shot problem is extended to cross-domain few-shot problem,the existing meta-learning models will perform poorly due to insufficient cross-domain capabilities.The cross-domain meta-learning approach improves on the meta-learning model by combining the cross-domain approach,so that the model can learn meta-knowledge with transferability,and the generalization of the meta-learning model can be improved.There are two settings of cross-domain meta-learning,namely,cross-domain meta-learning with target domain unavailable and target domain partially available.The former approach suffers from the lack of diversity and robustness of the model extracted features,and the latter approach suffers from the problem that the model is not biased enough to the target domain.To address the above problems,the specific research contents for the cross-domain few-shot classification problem in this paper are as follows:(1)To improve the diversity of features extracted by the model,a cross-domain metalearning model based on Self-Challenging Mask is proposed.We use the spatial activation map output from the cross-level discriminator to generate a Self-Challenging Mask for midlevel feature map,which is used to weaken the locations where the model focuses on and force the model to discover the features in other locations.We apply Self-Challenging Mask to all samples in the few-shot task and further extract to the high-level features,generating a challenging task to train the model,thus increasing the diversity of features extracted by the model.(2)To improve the robustness of the model for extracting features,a cross-domain meta-learning model based on Task-Aware Adversarial Feature Perturbation is proposed.The model consists of three components:Adversarial Feature Perturbation,Task Attention Module,and the regularization loss.Among them,Adversarial Feature Perturbation is the adversarial perturbation attack on each block of the backbone network.Task Attention Module is the feature encoding module that uses the sample relationship information in the task to exploits the discriminative features.The two complement each other so that the perturbation contains sample relationship information and also prompts the model to extract more robust discriminative features.The regularization loss requires the model to remain semantically invariant under adversarial perturbations and further improves the model robustness.(3)To improve the domain adaptability of the model to the target domain,a crossdomain meta-learning model based on Feature Meta-Augumentation Module is proposed.The model belongs to the cross-domain meta-learning method with target domain partially available.The Feature Meta-Augumentation Module is an mask-augumentation module for mid-level feature map.We make use of the idea of optimal meta-learning to meta-train the Feature Meta-Augumentation Module with a small number of samples from the target domain to make the model generalize well on the target domain.