Abstract

Generative Adversarial Networks (GAN) boast impressive capacity to generaterealistic images. However, like much of the field of deep learning, theyrequire an inordinate amount of data to produce results, thereby limiting theirusefulness in generating novelty. In the same vein, recent advances inmeta-learning have opened the door to many few-shot learning applications. Inthe present work, we propose Few-shot Image Generation using Reptile (FIGR), aGAN meta-trained with Reptile. Our model successfully generates novel images onboth MNIST and Omniglot with as little as 4 images from an unseen class. Wefurther contribute FIGR-8, a new dataset for few-shot image generation, whichcontains 1,548,944 icons categorized in over 18,409 classes. Trained on FIGR-8,initial results show that our model can generalize to more advanced concepts(such as "bird" and "knife") from as few as 8 samples from a previously unseenclass of images and as little as 10 training steps through those 8 images. Thiswork demonstrates the potential of training a GAN for few-shot image generationand aims to set a new benchmark for future work in the domain.

 

Introduction (beta)

None

 

Conclusion (beta)

None