OmnImage: Evolving 1k Image Cliques for Few-Shot Learning

Abstract

Few-shot learning datasets contain a large number of classes with only a few examples in each. Existing datasets may contain thousands of classes, but very simple images (e.g. handwritten characters) such that a naive baseline can perform very well. Or they may be so complex, with large within-class variation and distracting background, that they are too difficult to enable meaningful learning with so few examples. To construct a customizable dataset of consistent natural images, we assemble a new dataset with each class containing a small subset from each of the 1000 classes of ImageNet-1k. To select subsets with clear within-class consistency we use an evolutionary approach to minimize the pairwise cosine-distance between features generated by a pre-trained VGG model. We train a classifier on these evolved image cliques and find that our evolved dataset provides a greater challenge than hand written digits, but not the extreme difficulty of a non-evolved subset of ImageNet. We find that pre-training our classifier on these evolved prototypical classes significantly improves performance on classifying random subsets ImageNet (relative to pre-training on similar random-class subsets), and conjecture that these prototypical classes may be beneficial for seeding concept learning. Dataset and code is publicly available at: https://github.com/lfrati/OmnImage