Hard-Positive Prototypical Networks for Few-Shot Classification

Prominent prototype-based classification (PbC) approaches, such as Prototypical Networks (ProtoNet), use the average of samples within a class as the class prototype. In these methods which we call Mean-PbC, a discriminant classifier is defined based on the minimum Mahalanobis distance from class prototypes. It is well known that if the data of each class is normally distributed, then the use of Mahalanobis distance leads to an optimal discriminant classifier. We propose the Hard-Positive Prototypical Networks (HPP-Net), which also employs the Mahalanobis distance, despite assuming the class distribution may be unnormalized. HPP-Net learns class prototypes from hard (near-boundary) samples that are less similar to the class center and have a higher misclassification probability. It also employs a learnable parameter to capture the covariance of samples around the new prototypes. The valuable finding of this paper is that a more accurate discriminant classifier can be attained by applying the Mahalanobis distance in which the mean is a "hard-positive prototype", and the covariance is learned via the model. The experimental results on Omniglot, CUB, miniImagenet and CIFAR-100 datasets demonstrate that HPP-Net achieves competitive performance compared to ProtoNet and several other prototype-based few-shot learning (FSL) methods.