| Classification is a common machine learning task,which can be divided into binary classification and multi-classification.It plays an important role in machine vision,intelligent security,smart city,and other fields.In the classification task,continuous data,numerous sources,and low-value density increase the difficulty of the application of the classification task.Based on the idea of conjugate gradient optimization,a machine learning classification method based on conjugate gradient is studied in this thesis.The main research contents are as follows:(1)A monotone conjugate gradient binary classification algorithm based on stochastic optimization is proposed.Aiming at the problem that the traditional batch learning algorithm takes a long time to deal with the large-scale binary classification problem,the conjugate gradient method based on random optimization is used to solve the classification model and accelerate the convergence speed of the algorithm.The algorithm solves the problem of slow convergence of model in large-scale data set by uniform random sampling and calculation of conjugate direction,alleviates the oscillations of machine learning classification algorithm in the late training period,and improves the classification effect of machine learning in the classification task.In theory,the linear convergence rate of the proposed algorithm is proved.The convergence speed and running time of the algorithm are compared on several benchmark classification data sets.The comparison results show that the monotone conjugate gradient dichotomy algorithm based on random optimization can obtain more sparse solutions and has a faster convergence speed.(2)A non-monotone conjugate gradient multi-classification algorithm based on feedforward neural network is proposed.Aiming at the problem that feedforward network is prone to zigzag phenomenon and fall into local optimal solution when dealing with multiple classification problems,a non-monotone conjugate gradient multi-classification algorithm based on feedforward neural network is proposed by adopting conjugate gradient descent method and non-monotone linear search technology.In the multi-classification task,the algorithm makes full use of the loss function and its gradient information,and has faster convergence speed and better generalization without increasing the memory footprint.By analyzing the convergence and accuracy of the proposed algorithm on six common data sets,it is proved that the proposed algorithm has good generalization ability. |
