Few-shot Learning with Data Augmentation with Generative Model.

Abstract

While deep learning, in general, requires a large amount of labeled data, there are situations where only a few samples are available for some classes. In theory, if we can predict the probabilistic distribution of the classes based on the samples for other classes, we can leverage the distribution to train the model. We augment the data for the class with few samples using the generative model trained on the other classes for a classification task. We applied this method on MNIST dataset and evaluate it.