Research On Adaptive Manifold Meta-Learning

Meta-learning,also known as Learning to learn,refers to the use ofprevious knowledge and experience to guide the learning of new tasks.Although meta-learning has made great progress in the field of image recognition,but there are still many problems to be solved,such as adaptive problem,generalization problem,interpretability problem and so on.This thesis mainly aim at the adaptive problems of meta-learning in this field,from three aspects of the self-adaptive sample’s weight,self-adaptive sample’s distribution,self-adaptive feature measurement to conduct research.At the same time,in order to alleviate the overfitting problem,manifold learning is added to this thesis.Based on the prototype network,the research results obtained in this thesis include:(1)An daptive sample’s weight manifold meta-learning algorithm(ASWMM)is proposed.The existing meta-learning model cannot distinguish between different samples effectively.The ASWMM algorithm add the adaptive method makes the model distinguish samples of different quality.The algorithm includes an adaptive sample weight module,each sample is given a hyperparameter to measure the sample quality.At the same time,the algorithm also includes a sample difference enlargement formula to further expand the differences between samples.Finally,the validity of the algorithm is proved on common datasets(Ominglot,Minilmagenet,CUB).(2)An adaptive sample’s distribution manifold meta-learning algorithm(ASDMM)is proposed.The ASWMM algorithm focuses on each sample,but this algorithm treats all samples as a whole from a macro perspective.When the distribution of samples in multidimensional space is unreasonable,the common meta-learning algorithms can not make effective tune.The adaptive sample distribution manifold learning algorithm proposed in this thesis can adaptively adjust the sample’s distribution in high dimensional space.The algorithm consists of two parts,respectively an aggregation module where similar samples are gathered together and a discrete module where non-similar samples are separated.Finally the validity of the algorithm is proved in common datasets(Ominglot,MiniImagenet,CUB).(3)An adaptive feature metric manifold learning algorithm(AFMMM)is proposed.ASWMM and ASDMM algorithms are studied from the perspective of samples,rather than detailed studies specific characteristics of the sample.This algorithm mainly studies the characteristics of samples in more detail.In high-dimensional Spaces,the importance of the features extracted by the model is not the same.The algorithm introduces two adaptive measurement methods to measure the importance of sample features.Finally,the validity of the algorithm is proved in common datasets(Ominglot,MiniImagenet,CUB).