Missing Glow Phenomenon: learning disentangled representation of missing data

Anonymous

Keywords: deep learning, disentanglement, invertible models, flow models, Glow, missing data

TL;DR: This paper shows that flow-based generative models like Glow (Kingma & Dhariwal, 2018) can work directly on images with missing data to produce full images without missing parts. We name this behavior Missing Glow Phenomenon.

Abstract:

Learning from incomplete data has been recognized as one of the fundamental challenges in deep learning. There are many more or less complicated methods for processing missing data by neural networks in the literature.
In this paper, we show that flow-based generative models can work directly on images with missing data to produce full images without missing parts. We name this behavior Missing Glow Phenomenon. We present experiments that document such behaviors and propose theoretical justification of such phenomena.