Enlarge the Hidden Distance: A More Distinctive Embedding to Tell Apart Unknowns for Few-Shot Learning.

Most few-shot classifiers assume consistency of the training and testing distributions. However, in many practical applications, the two distributions are often different. In this paper, we focus on the few-shot open-set recognition problem which allows that the testing categories are different from the training categories. To alleviate this problem, we take the semantic adhesion scenario as an example to analyze the influence of sample embedding vectors on the identification indicator value. Then, we propose an Extra Embedding Classification Model with an adjustment module that is trained with optional contrastive loss functions to learn distinctive features of samples in the same category. This model can enlarge the hidden distances among samples while keeping the category information. We comprehensively verified the effectiveness of our model on both the normal and the semantic adhesion scenario of the few-shot open-set recognition problem.