Linear subspace networks for few-shot classification

Abstract

We address the limitation of prototypical network, which is a popular metric-based few-shot learning algorithm. Prototypical networks use the center of embedding features as prototypes to alleviate over-fitting problem in few-shot learning and are built on the idea that examples in the same class are be clustered at a point in the embedding space. However, few researchers have addressed the limitation of one-point prototypes and showed performance improvement by using multiple-point prototypes for each class. Especially, subspace networks learn an embedding function under the idea that examples in the same class are be clustered near a subspace in the embedding space. However, there is no discussion about reasons for using subspace prototypes instead of one-point prototypes. In this paper, we bridge the gap between prototypical networks and subspace networks by adding simple constraints, affine and directional constraint, on the ways to determine prototypes. If we add both constraints to subspace networks, then it is exactly the same as prototypical networks with Euclidean similarity. Consequently, we propose affine space networks and linear subspace networks by adding the affine and directional constraint to subspace network respectively. The experimental results of this study indicate that the directional constraint mainly is contributed to the performance improvement in few-shot classification tasks, although the affine constraint is not effective. Moreover, we propose a parametrized metric that generalizes both the prototypical network and linear subspace network. By using the trainable metric instead of Euclidean or cosine similarity for prototypical networks, the performance is quite high compared to the existing metric-based few-shot learning algorithms.