Selected Publications
SCAN: Learning Abstract Hierarchical Compositional Visual Concepts
This paper describes SCAN (Symbol-Concept Association Network), a new framework for learning recombinable concepts in the visual domain. We first use the previously published beta-VAE (Higgins et al., 2017a) architecture to learn a disentangled representation of the latent structure of the visual world, before training SCAN to extract abstract concepts grounded in such disentangled visual primitives through fast symbol association.
arXiv
dSprites: Disentanglement testing Sprites dataset
This dataset consists of 737,280 images of 2D shapes, procedurally generated from 5 ground truth independent latent factors, controlling the shape, scale, rotation and position of a sprite. This data can be used to assess the disentanglement properties of unsupervised learning methods.
Github