Few-shot Learning via Non-negative Representation

Due to the uncertainty caused by using a small number of labeled samples, few-shot classification is a challenging problem. In the past few years, many methods have been proposed to solve few-shot classification, among which the method based on transduction has been proved to be the best. According to this idea, in this paper, we propose a new transduction-based method, which is based on the nonnegative representation. In order to obtain the classification results, we minimize the objective function with two items: (1) The first item assigns representation coefficients for each class prototype to obtain the minimum reconstruction error; (2) The second item encourages similar query samples to have consistent label allocation. At the same time, we make non-negative constraints on the representation coefficient to make the representation sparse and discriminative. Using standardized visual benchmarks, we prove that the proposed method can achieve high accuracy in various data sets, inductive and transductive settings, and is robust to the situation that the number of unlabeled samples per class is unbalanced.