Few-Shot Learning with Variational Homoencoders


Nov 07, 2017 (modified: Nov 07, 2017) ICLR 2018 Conference Blind Submission readers: everyone Show Bibtex
  • Abstract: Few-shot learning of generative models typically falls into two philosophies: conditional models, trained to generate new elements conditioned on observations, and hierarchical Bayesian models, which frame conditioning as inference of latent variables shared by a class. We modify the Variational Autoencoder framework to marry these two approaches, learning a hierarchical Bayesian model by a novel training procedure in which observations are encoded and decoded into new elements from the same class. We call this a Variational Homoencoder (VHE) and apply it to Caltech 101 Silhouettes and the Omniglot dataset. On Omniglot, our hierarchical PixelCNN outperforms existing models on joint likelihood of the data set, and achieves state-of-the-art results on both one-shot generation and one-shot classification tasks. The VHE framework also applies naturally to richer latent structures such as factorial or hierarchical categories. We illustrate this by training models to separate character content from simple variations in drawing style, and to generalise the style of an alphabet to new characters.
  • TL;DR: Technique for learning deep generative models with shared latent variables, applied to Omniglot with a PixelCNN decoder.
  • Keywords: generative models, one-shot learning, metalearning, pixelcnn, hierarchical bayesian, omniglot