BEpiC: Binary Episodes for Meta-Learning Towards Better Generalization

In practical machine learning applications, image classification often involves discerning the category to which an object belongs. In this paper, we present a novel approach called Binary Episode Classifier (BEpiC) within the context of meta-learning. BEpiC aims to optimize episode generation by strategically selecting training samples. The methodology involves training the initial class with a comprehensive set of similar images while the contrasting class is exposed to a diverse array of dissimilar images. We create highly dissimilar image sets by randomly forming multiple image clusters. Identifying the cluster centers and selecting the representative image closest to each center facilitates the determination of the among clusters. The distance of these images from the other class is then calculated. The cluster whose representative image exhibits the greatest distance is chosen as the definitive class. Our proposed method showcases significant efficacy for binary meta-learning classification across diverse classes. While our experimentation primarily focused on Model-Agnostic Meta-Learning (MAML), the adaptability of the episode generation strategy extends to a spectrum of meta-learning classifiers. Empirical findings substantiate that the proposed method attains remarkable accuracy in one-shot classification scenarios and moderately higher accuracy in few-shot classification tasks.