Denoising Multi-Similarity Formulation: A Self-Paced Curriculum-Driven Approach for Robust Metric Learning

Deep Metric Learning (DML) is a group of techniques that aim to measure the similarity between objects through neural network. Although number DML methods has rapidly increased in recent years, most previous studies cannot effectively handle noisy data, which commonly exists practical applications and often leads serious performance deterioration. To overcome this limitation, paper, we build connection samples hard framework self-paced learning, propose Balanced Self-Paced (BSPML) algorithm with denoising multi-similarity formulation, where are treated as extremely adaptively excluded from model training by sample weighting. Especially, due pairwise relationship new balance regularization term, sub-problem w.r.t. weights nonconvex quadratic function. efficiently solve problem, doubly stochastic projection coordinate gradient algorithm. Importantly, theoretically prove convergence not only for algorithm, but also our BSPML Experimental results on several standard data sets demonstrate better generalization ability robustness than state-of-the-art robust approaches.