Distance Metric Optimization-Driven Neural Network Learning Framework for Pattern Classification

As a novel neural network learning framework, Twin Extreme Learning Machine (TELM) has received extensive attention and research in the field of machine learning. However, TELM is affected by noise or outliers practical applications so that its generalization performance reduced compared to robust algorithms. In this paper, we propose two distance metric optimization-driven twin extreme frameworks for pattern classification, namely, CWTELM FCWTELM. By introducing Welsch loss function capped L2,p-distance metric, our methods reduce effect improve model TELM. addition, efficient iterative algorithms are designed solve challenges brought non-convex optimization problems FCWTELM, theoretically guarantee their convergence, local optimality, computational complexity. Then, proposed with five other classical under different datasets, statistical detection analysis implemented. Finally, conclude algorithm excellent robustness classification performance.