Research On Few-Shot Learning Method Based On Meta-Learning

Few-shot learning refers to learning an effective classifier from a dataset with very few samples.Deep learning typically requires a large amount of data for training,but in reality,we often only have limited data,and deep learning performs poorly in handling few-shot problems.The main challenges faced include: 1)Limited data that cannot meet the requirements of deep learning,making it difficult for the model to learn effectively.2)Traditional transfer learning methods are also difficult to handle few-shot problems.Because in few-shot learning,there is often very little overlap between different tasks,and the model cannot learn enough knowledge from existing data.3)In few-shot learning,the model needs to be trained on very little data.Therefore,the complexity of the model is strictly limited,and if the model is too complex,overfitting is likely to occur.Aim at above problems,this paper proposes a research framework that integrates attention modules into the meta-learning process.By exploiting more information from the data itself through attention modules,the framework adds more supervisory signals for training models from the perspective of information gain.In this paper,a few-shot learning method based on global class representation is first proposed,which optimizes the global class representation from the perspective of data augmentation,improving the model’s generalization ability to new class data.Furthermore,in order to fully utilize the features of unlabeled samples,a few-shot learning method based on interactive attention is proposed,which models the semantic correlation between features through interactive attention to enhance inter-class separability.The main work of the paper includes:First,a few-shot learning method based on global class representation with selfattention mechanism is proposed.Firstly,a small amount of new class data is added to the training phase,and a global class representation is constructed for all categories using a pretrained feature extraction network.Then,in the meta-task process,after data augmentation of the new class data,it is input together with the base class data into the feature extraction network to match a global class representation for each local class.Finally,the feature maps of the samples in the query set are input together with the global class representation into the attention network for category prediction.This method has been effectively validated on the Omniglot and miniImageNet datasets.Second,a few-shot learning method of global class representation based on interactive attention is proposed.Firstly,the training samples are expanded by direct learning to improve the feature richness of categories.Then,in the meta-task stage,the cross-attention module is introduced to construct the semantic correlation between class features and query features,so as to attract the attention of the target object to achieve the prediction of query samples.The experimental results show that the simultaneous introduction of direct learning and cross-attention module has a significant improvement in the classification task.Thirdly,a few-shot data classification system based on meta-learning is designed and completed.The system has five modules,which can improve the efficiency of algorithm parameter adjustment and conveniently and intuitively display the comparison results of different algorithms on different data sets.