Imagining the Unseen: Learning a Distribution over Incomplete Images with Dense Latent Trees

• DOI: 10.48550/arxiv.1808.04745
• Publication Date: 2018-01-01
• Authors: Sebastian Kaltwang; Sina Samangooei; John Redford; Andrew Blake
• Keywords: MNIST database; Generative model; Graphical model; Hierarchical database model; Tree (set theory)

Images are composed as a hierarchy of object parts. We use this insight to create generative graphical model that defines hierarchical distribution over image Typically, leads intractable inference due loops in the graph. propose an alternative structure, Dense Latent Tree (DLT), which avoids and allows for efficient exact inference, while maintaining dense connectivity between parts hierarchy. The usefulness DLTs is shown example task completion on partially observed MNIST Fashion-MNIST data. verify having successfully learned images by visualising its latent states.