Transformation-Based Data Synthesis for Limited Sample Scenario

Abstract

We consider a challenging learning scenario where neither pretext training nor auxiliary data are available except for small training samples. We call this a transfer-free scenario where we cannot access any transferable knowledge or data. Our proposal for resolving this issue is to learn a pair-wise transformation function (e.g., spatial or appearance) between given samples. This simple setting yields two practical advantages. The training objective can be defined as a simple reconstruction loss, and data can be synthesized by merely manipulating or sampling the learned transformations. However, the limitation of previous transformation methods lies in a strong assumption that all images should be transformable to each other, i.e., all-to-all transformable. To relax this constraint, we propose a novel concept called 'template,' designed to be transformable to any other data, i.e., "template-to-all" transformable. A range of experiments on the transfer-free scenarios confirms that our model successfully learns transformation and synthesizes new data from minimal training data (less than five or ten for each class). The subsequent data augmentation experiments showed significantly improved classification performance.