Building Models that Learn to Discover Structure and Relations

Some argue that a key component of human intelligence is our ability to reason about objects and their relations (e.g. This enables us, for example, to build rich compositional models of physics (how objects or particles interact) and intuitive theories of causation (what causes what) [3]. For artificial systems, these tasks remain a challenge. Most sophisticated pattern recognition models, e.g. based on Convolutional Neural Networks (CNNs) or Recurrent Neural Networks (RNNs), lack a certain relational inductive bias [4]; impeding their ability to generalize well on problems with inherent compositional structure. In our recent ICML (2018) paper: Neural Relational Inference for Interacting Systems, we explore a class of models named Graph Neural Networks (GNNs) that reflect the inherent structure of the problem domain in their model architecture¹.