Dynamic Broad Metric Learning

Many existing metric learning methods are based on fixed similarity constraints. However, the quality of fixed similarity constraints is usually hard to guarantee, and inflexible constraints also limit the performance of metric learning. Moreover, when new training samples are acquired, retraining the model is usually inefficient for many classical metric learning methods. This article proposes a novel dynamic broad metric learning (DynBML) method to overcome the above limitations. DynBML trains a broad network with dynamic similarity constraints between samples and their target points to learn a discriminative data transformation with high representation ability. To overcome the weaknesses of fixed constraints, DynBML dynamically updates the target points to produce new constraints during the learning process. A force-directed reference points generation (FD-RPG) algorithm is designed to generate reference points from the currently learned space, which encourages the new target points to be more discriminative. A new objective function is proposed to simultaneously learn the new target points and train the data transformation with the guidance of the reference points. Additionally, to avoid retraining the model from scratch when new data arrive, the incremental learning scheme of DynBML (I-DynBML) is developed, which efficiently updates the trained model of DynBML by generating and satisfying new constraints of new input. DynBML is evaluated on ten datasets to demonstrate its superiority. Experiment results show that the dynamic constraints constructed in DynBML improve the learning performance. The effectiveness and efficiency of the incremental learning scheme of DynBML are also verified.