One-Class Meta-Learning: Towards Generalizable Few-Shot Open-Set Classification

  • 2021-09-14 17:52:51
  • Jedrzej Kozerawski, Matthew Turk
  • 1


Real-world classification tasks are frequently required to work in anopen-set setting. This is especially challenging for few-shot learning problemsdue to the small sample size for each known category, which prevents existingopen-set methods from working effectively; however, most multiclass few-shotmethods are limited to closed-set scenarios. In this work, we address theproblem of few-shot open-set classification by first proposing methods forfew-shot one-class classification and then extending them to few-shotmulticlass open-set classification. We introduce two independent few-shotone-class classification methods: Meta Binary Cross-Entropy (Meta-BCE), whichlearns a separate feature representation for one-class classification, andOne-Class Meta-Learning (OCML), which learns to generate one-class classifiersgiven standard multiclass feature representation. Both methods can augment anyexisting few-shot learning method without requiring retraining to work in afew-shot multiclass open-set setting without degrading its closed-setperformance. We demonstrate the benefits and drawbacks of both methods indifferent problem settings and evaluate them on three standard benchmarkdatasets, miniImageNet, tieredImageNet, and Caltech-UCSD-Birds-200-2011, wherethey surpass the state-of-the-art methods in the few-shot multiclass open-setand few-shot one-class tasks.


Quick Read (beta)

loading the full paper ...