Research On The Improvement Of Metric Learning Algorithm Based On Deep Neural Networks

In the research of machine learning,the appropriate metric algorithms are often used to calculate the distance or similarity between samples.Metric learning can learn automatically according to the characteristics of the data,which not only provides a rich theoretical tool for other algorithms but also can solve practical problems.However,for the existing metric learning algorithms,there are some shortcomings in model optimization and model generalization.In view of this,the thesis studies the improved algorithm of metric learning based on neural network.The thesis first introduces the basic concepts of metric learning,analyzes the complexity of different constraint construction strategies and summarizes the existing regularization terms and loss functions.Furthermore,the thesis analyzes the relationship between the number of samples in the loss function and its performance,and summarizes the model structure of metric learning.In order to solve the difficulty in communication between single-layer metric learning and deep metric learning optimization technology,this thesis proposes a hierarchical optimization metric learning algorithm based on neural network.The algorithm consists of two levels: the first level extracts the key features from the original data with the help of the pre-network,and the second level maps the key features to the embedded space with the help of the projection matrix.The algorithm utilizes the back propagation algorithm and strict geodesic convex optimization technique to optimize the parameters,respectively,and designs the regularization terms on the basis of cosine similarity to avoid the key homogeneous characteristics of the front network.Furthermore,this paper conducts simulation experiments using standard test data sets and handwritten digital data set.Compared with other classical algorithms in K nearest neighbor classification error rate and visual renderings,the proposed algorithm can better realize data reduction and improve data distribution.For the problem of unbalanced data distribution in the existing algorithms,a normalized metric learning based on deep Fractal Net is proposed.The fractal network designed by the algorithm consists of two branches.One branch normalizes the same kind of samples to the standard normal distribution,and the other branch calculates the mean vector of each category.The algorithm utilizes KL divergence and Monte Carlo method to design the loss function,to calculate the mean and variance of the normal distribution and the difference between the data distribution.Furthermore,this paper conducts simulation experiments using CUB200-201 and CARS-196 data set.Compared with the existing algorithms in standardized mutual information and recall@K,the proposed algorithm can effectively avoid the problem of unbalanced distribution and has better generalization ability.