Research On Meta-Learning Algorithms Oriented To Tasks With Multi-Distribution

With the rapid development of machine learning,the ability of machine learning model has been continuously strengthened in recent years,and even surpass the mankind in some fields.However,the training process of most existing machine learning models relies on a large number of manually labeled training samples,while humans are able to learn new skill quickly with a few of training data.Meta-learning aims to accelerate the learning of new tasks by reusing the prior knowledge stored in the previously learned tasks.The improvement of model training based on the meta-learning algorithm can effectively reduce the training cost of the model under new tasks.In this thesis,In this paper,we conduct research on the learning and optimization of meta-learning algorithms in tasks with multidistribution.The main research results are as follows:(1)Aiming at the problems of unreasonable storage and wrong reuse of prior knowledge on the meta-training stage in existing meta-learning algorithms.An online clusteringbased meta-learing algorithm is proposed in this paper.The algorithm is devoted to integrating meta-training tasks from different task distributions into divergent clustering features through online clustering,and selectively reusing the most relevant clustering feature for model parameter initialization through the minimum distance matching algorithm in the meta-testing stage.Furthermore,the algorithm also designs an evaluation network to generate the correlation coefficient between the current task feature and the selected cluster feature,which helps the model to obtain more reliable initial model parameter for the current task.At last,a series of few-shot classification experiments are designed to verify the performance of the proposed algorithm,experiments results reveal that the proposed algorithm can reasonably store and reuse relevant prior knowledge to accelerate the learning of new tasks.(2)Aiming at the problem of uncertain distribution of meta-training tasks in metalearning scenarios,this paper proposes a meta-learning algorithm based on online clustering with selective rollback.The proposed algorithm can automatically increase and update the relevant clustering features by the way of selectively rollback during model meta-training,and efficiently reuse the relevant prior knowledge during meta-testing stage.In addition,for the lack of permutation invariance of task features extracted in the task feature extraction stage of the current algorithm,the algorithm consider to use DeepSets as the backbone network for the task feature extraction layer,for that the model can generate task feature with permutation invariance.A series of experiments show that the algorithm can effectively execute task feature clustering in the scenario where the number of task distribution is unknown.At the same time,the DeepSets used in the feature extraction layer can also help the model to extract task features more efficiently.