TAFSSL: Task-Adaptive Feature Sub-Space Learning for Few-Shot Classification

Moshe Lichtenstein, Prasanna Sattigeri, Rogério Feris, Raja Giryes, Leonid Karlinsky

Abstract: Recently, Few-Shot Learning (FSL), or learning from very few (typically 1 or 5) examples per novel class (unseen during training), has received a lot of attention and significant performance advances. While number of techniques have been proposed for FSL, several factors have emerged as most important for FSL performance, awarding SOTA even to the simplest of techniques. These are: the backbone architecture (bigger is better), type of pre-training (meta-training vs multi-class), quantity and diversity of the base classes (the more the merrier), and using auxiliary self-supervised tasks (a proxy for increasing the diversity). In this paper we propose TAFSSL, a simple technique for improving the few shot performance in cases when some additional unlabeled data accompanies the few-shot task. TAFSSL is built upon the intuition of reducing the feature and sampling noise inherent to few-shot tasks comprised of novel classes unseen during pre-training. Specifically, we show that on the challenging miniImageNet and tieredImageNet benchmarks, TAFSSL can improve the current state-of-the-art in both transductive and semi-supervised FSL settings by more than $$5\%$$ , while increasing the benefit of using unlabeled data in FSL to above $$10\%$$ performance gain.