Behavior regularized prototypical networks for semi-supervised few-shot image classification

We propose a Behavior Regularized Prototypical Network (BR-ProtoNet) for few-shot image classification in semi-supervised scenarios. To learn a generalizable metric, we exploit readily-available unlabeled data and construct complementary constraints to regularize the model's behavior. Specifically, we match the label spaces between each episode and the whole training set. The predictions on the unlabeled data over different episodes can be aggregated to capture more reliable category information. We further construct new instances via adversarial perturbation and interpolation. These instances regularize the model's behavior over the neighborhoods of the original ones and along the interpolation paths among them. In addition, they ensure the learnt embedding space possesses the property of proximity preservation. The regularization of these aspects is incorporated into the optimization process of BR-ProtoNet on partially labeled data. We have conducted thorough experiments on multiple challenging benchmarks. The results suggest that the metric learning can significantly benefit from the proposed regularization, and thus leading to the state-of-the-art performance in semi-supervised few-shot image classification.