Object-Oriented Learning (OOL): Perception, Representation, and Reasoning
International Conference on Machine Learning (ICML)
Friday July 17, 2020, Virtual Workshop
One of the ultimate goals of unsupervised representation learning is identifiability, or the ability to recover the true generating factors for a set of observations. For object-oriented representations, learning an identifiable representation for the constituent objects in a scene would mean that we have effectively learned each entity's true latent code. Recent breakthroughs in nonlinear ICA have shown how identifiability can be achieved in deep latent variable models. Currently, however, no image datasets exist which allow for quantitatively evaluating the performance of models based on the proposed theories. To this end, we introduce iSprites, a dataset for identifiable single- and multi-object representation learning. Contrary to previously published datasets, iSprites facilitates the testing of representation learning methods which are theoretically identifiable, thus providing a bridge between recent theory and practical applicability.