Research On Metric Adaptive Meta-learning

As a machine learning method,meta-learning has achieved remarkable results in many fields.However,compared with deep learning,the current meta-learning method still has considerable room for improvement,and this paper mainly studies the application of metalearning in supervised learning from the perspective of feature and distance measurement,and proposes an overall framework of metric adaptive meta-learning algorithm.The main work of this paper includes:(1)Adaptive metric parameters Network for Few-shot Learning(AMPA)is proposed to adapt the metric method to the data distribution characteristics of the task in terms of spatial distribution of data.AMPA addresses the issue of poor adaptability to data distribution in metric based meta learning models by adding a representation of the distribution characteristics of the current task data,correcting the weights and data biases of measurement methods,enabling different types of data in the feature space to be correctly distinguished and narrowing the distribution differences between support and query sets.(2)Partial feature metric algorithm for meta-learning(PFMA)is proposed to change the computational proportion of dominant features and Partial features at the feature level.Common meta learning algorithms overlook the Partial features of sample data,resulting in a decrease in the algorithm’s ability to distinguish between different types of nearest neighbors.During the training process,PFMA learns the existing explicit features of the sample,generates and completes the implicit features of the sample,and enhances the weight of the implicit features at the feature level,thereby improving the accuracy of sample point recognition.(3)Subspace mapping metric algorithm for meta-learning(SSMA)is proposed.The task space is decomposed into singular values,and the task space is mapped into orthogonal subspaces using a left singular matrix.Based on the singular values,the feature dimensions are filtered and reduced,reducing the interference of noise features,and improving the accuracy of Euclidean metric methods for tasks in Euclidean space.