Adaptive Hypersphere Data Description for few-shot one-class classification

Abstract

Few-shot one-class classification (FS-OCC) is an important and challenging problem involving the recognition of a class using a limited number of positive training samples. Data description is essential for solving the FS-OCC problem as it delineates a region that separates positive data from other classes in the feature space. This paper introduces an effective FS-OCC model named Adaptive Hypersphere Data Description (AHDD). AHDD utilizes hypersphere-based data description with a learnable radius to determine the appropriate region for positive samples in the feature space. Both the radius and the feature network are learned concurrently using meta-learning. We propose a loss function for AHDD that enables the mutual adaptation of the radius and feature within a single FS-OCC task. AHDD significantly outperforms other state-of-the-art FS-OCC methods across various benchmarks and demonstrates strong performance on test sets with extreme class imbalance rates. Experimental results indicate that AHDD learns a robust feature representation, and the implementation of an adaptive radius can also improve the existing FS-OCC baselines.