Meta-learning frameworks are usually trained in a few-shot setting and suffer from data scarcity. A straightforward solution to this problem is using data augmentation to expand our support set used for training the meta-learning parameters. But applying the same augmentation for all input instances can be problematic. For instance, applying the same 180-degree rotation to a dataset of images may change a 6 to a 9, or applying the same color jitter may change a lemon to a lime. While prior works in data augmentation have largely focused on instance-agnostic augmentation strategies, we propose a novel technique for generating instance-based augmentations called AutoInstanceAug which is adapted for meta-learning setup where we don’t have sufficient data to train.
