Semi-Supervised Metric Learning: A Deep Resurrection

Distance Metric Learning (DML) seeks to learn a discriminative embedding where similar examples are closer, and dissimilar apart. In this paper, we address the problem of Semi-Supervised DML (SSDML) that tries metric using few labeled examples, abundantly available unlabeled examples. SSDML is important because it infeasible manually annotate all present in large dataset. Surprisingly, with exception classical approaches linear Mahalanobis metric, has not been studied recent years, lacks deep scenario. challenging problem, revamp respect learning. particular, propose stochastic, graph-based approach first propagates affinities between pairs from data, pairs. The propagated used mine triplet based constraints for We impose orthogonality constraint on parameters, as leads better performance by avoiding model collapse.