Research On Noise Robust Algorithm Based On Self-paced Learnin

In recent years,machine learning has played an important role in more and more practical applications.However,due to the presence of noisy samples,the performance of machine learning algorithms is often affected.Self-paced learning is a training strategy that learns from easy to difficult,and it simulates the learning process of humans and animals.Numerous experimental results and theoretical analysis have shown that self-paced learning can improve the noise-robustness of the model through sample weighting.AUC(Area Under the Curve)optimization and deep metric learning are important research topics in the field of machine learning.AUC optimization is widely used in scenarios with unbalanced categories,such as information retrieval,medical diagnosis,and fault detection.Technologies related to deep metric learning have brought many conveniences to people’s lives,such as visual tracking,face recognition,and image retrieval.However,existing AUC optimization algorithms and deep metric learning algorithms cannot effectively process noisy samples.In view of this,this paper uses self-paced learning to improve the noise-robustness of the above two algorithms:(1)This paper proposes a noise-robust AUC optimization algorithm based on self-paced learning,namely the balanced self-paced AUC optimization(BSPAUC)algorithm.Unlike traditional self-paced learning algorithms,the BSPAUC algorithm needs to consider more subproblems,i.e.,sub-problems with respect to positive sample weights,negative sample weights,and model parameters.To overcome this challenge,this paper proposes a doubly cyclic block coordinate descent method.In order to efficiently solve the sub-problem with respect to sample weights,this paper provides analytical solutions for positive and negative sample weights through theoretical analysis.Experimental results show that compared to other AUC optimization algorithms,the BSPAUC algorithm has significant advantages in resisting noisy samples.(2)This paper proposes a noise-robust deep metric learning algorithm based on self-paced learning,namely the balanced self-paced deep metric learning(BSPML)algorithm.Specifically,in order to avoid the situation where a certain category is ignored under the self-paced learning strategy,this paper proposes a multi-classification balanced regularization term.Like traditional self-paced learning algorithms,the BSPML algorithm iteratively solves sub-problems with respect to sample weights and model parameters.The difference is that the sub-problem with respect to sample weights in the BSPML algorithm is a non-convex quadratic problem that is difficult to solve.To efficiently solve this problem,this paper proposes a doubly stochastic projection coordinate gradient method.Experimental results show that the BSPML algorithm has stronger noise-robustness compared to other deep metric learning algorithms.