Object-Oriented Learning (OOL): Perception, Representation, and Reasoning
International Conference on Machine Learning (ICML)
Friday July 17, 2020, Virtual Workshop
Recently, there has been a growing interest in incorporating relational reasoning into neural networks. However, existing approaches typically extract objects from inputs through an arbitrary partitioning of the latent space. We demonstrate that these methods do not respect set invariance and thus have fundamental representational limitations. We propose a novel and general network module that resolves this limitation and demonstrate how it naturally induces object decomposition without explicit supervision. As part of this, we insert our module into existing relational reasoning models and see significant performance gains by respecting set invariance.