A Study Of Few-shot Learning Based On Knowledge Prior And Task Augmentation

Most of the current research in the field of machine learning focuses on training on large-scale dataset.However,in real-world scenarios,traditional deep learning models have insufficient generalization ability with limited data,lagging far behind human performance.How to learn with limited data is a prominent issue of concern,and fewshot learning has emerged to tackle this challenge.There are three main categories of few shot learning methods:data augmentation,transfer learning and meta-learning.Data augmentation methods essentially hope to expand data to solve few shot problems,which have limited gain in the case of extremely small data.Transfer learning learns relevant knowledge by training on a source domain with abundant data and transferring the knowledge to a target domain with relatively less data.However,the performance of transfer learning depends on the correlation between the source and target domains,and it does not address the issue of poor generalization ability in few-shot learning.Methods based on meta-learning have shown relatively good performance on multiple few shot tasks,and current investigations on few shot tasks are mainly focused on the field of meta-learning.In this paper,we further explored few shot learning based on meta-learning.Compared to traditional deep learning,the meta-learning-based approach does perform well on various few shot tasks,but the model performance is still poor for extremely small samples(less than 50 items in a single category),and the model is sensitive to data due to the relatively small amount of annotated data,and tends to over-fit a small amount of data,with problems such as poor robustness and stability.In order to effectively improve the model performance and alleviate the problems of poor robustness and stability of the model due to the relatively small amount of data.This paper proposes the exploitation of data and the prior knowledge to help further optimise the model and improve its performance.This paper explores three aspects of this.1.First,we improve the performance of few-shot event detection algorithms by injecting additional prior knowledge into the model.Based on the prototype network,we propose a few-shot event detection model using label augmentation and contrastive learning.By constructing templates to introduce external knowledge of labels,the model improves sample representation and reduces the distance between representations of samples from the same class.Then,we utilize contrastive learning to distance the samples from different classes,significantly improving the performance of few-shot event detection tasks.At the same time,the robustness and stability of the model are also significantly improved.2.Secondly,we improve the performance and effectiveness of few-shot event detection by designing a prefix generation network to mine task-specific prior knowledge.By freezing the parameters of the pre-trained model and fine-tuning the parameters of the prefix generation network,we generate prefix information corresponding to each layer of the pre-trained model for a single task,and uses the prefix to mine knowledge and improve sample representations,thereby improving the model performance.Experiments show that the prefix generation network can effectively mine internal knowledge of the data,optimize sample representations,and significantly improve the effectiveness of model parameters,while greatly saving the computational resources required for model training.3.Finally,we propose task selection and task augmentation to improve the performance of few-shot image recognition tasks.In the training phase,task-specific mixup augmentation is designed to generate a large number of new tasks to assist in model training.Then,by exploring the relationship between the gradient information and sample information of the training and testing tasks,training tasks with higher relevance to the testing task are selected for task augmentation,further improving the model performance.