Few-shot Classification with Hypersphere Modeling of Prototypes

Abstract

In few-shot classification, key points to make the learning phase effective are to construct expressive class-level representations and design appropriate metrics. However, previous studies often struggle to reconcile the expressivity of representations and the conciseness of metrics. When modeling class-level information, vanilla embeddings can make classification difficult due to the lack of capacity, whereas complex statistical modeling hinders metric interpretation. To address the issues simultaneously, this paper presents a simple and effective approach from the geometrical perspective, dubbed as hypersphere prototypes. Specifically, our method represents class information as hyperspheres, which are characterized by two sets of learnable parameters: a center and a radius. Our method enjoys the following advantages. (1) With the learnable parameters, unique class representations can be easily constructed and learned without additional restrictions. (2) Using “areas” instead of “points” as class representation, the expressive capability will be greatly enhanced, increasing the reliability of few-shot classification. (3) The metric design is intuitive for hypersphere representation, which is the distance from a data point to the surface of the hypersphere. As a fundamental method of few-shot classification, our method demonstrates remarkable effectiveness, generality, and compatibility with other technologies in experiments.