Clearing the confusion when learning to classify an image with a glance

What is it about?

Few-shot image classification is challenging due to the lack of ample samples in each class. Such a challenge becomes even tougher when the number of classes is very large, i.e., the large-class few-shot scenario. In this novel scenario, existing approaches do not perform well because they ignore confusable classes, namely similar classes that are difficult to distinguish from each other. These classes carry more information. We propose a biased learning paradigm called Confusable Learning, which focuses more on confusable classes. Our method can be applied to mainstream meta-learning algorithms.

Featured Image

Photo by Ilya Pavlov on Unsplash

Photo by Ilya Pavlov on Unsplash

Why is it important?

To tackle the large-class few-shot scenario, we propose a biased learning paradigm called Confusable Learning. Our key idea is to focus on confusable classes in meta-training dataset, which can improve model robustness in meta-testing dataset. In each iteration, we uniformly sample a few classes and denote them as target classes. For each target class, our paradigm selects several similar classes, which the model has difficulty in distinguishing from their target class. We call these classes distractorsFootnote 1. The model is then trained by a meta-learning algorithm to recognize instances of target class from those of distractors. Note that distractors are dynamically changing: when the model fits the distractors in each iteration, they become less confusable; while other classes become relatively more confusable and have higher chance to be selected as distractors. In this way, the model goes through every class in meta-training dataset dynamically.

Perspectives