Machine understanding of complex images is a key goal of artificial intelligence. One challenge underlying this task is that visual scenes contain multiple inter-related objects, and that global context plays an important role in interpreting the scene. A natural modeling framework for capturing such effects is structured prediction, which optimizes over complex labels, while modeling within-label interactions. However, it is unclear what principles should guide the design of a structured prediction model that utilizes the power of deep learning components. Here we propose a design principle for such architectures that follows from a natural requirement of permutation invariance. We prove a necessary and sufficient characterization for architectures that follow this invariance, and discuss its implication on model design. Finally, we show that the resulting model achieves new state of the art results on the Visual Genome scene graph labeling benchmark, outperforming all recent approaches.

This paper studies the property of permutation invariance in the context of structured prediction. The paper argues that in many applications permutation invariance is a desirable property of a solution and it makes sense to design the model such that it is satisfied by construction rather than to rely on learning to get this property. The paper proposes a model to represent permutation invariant functions and claims that this model is a universal approximator within this family. The proposed method is evaluated on a synthetic and a real task (labelling of scene graphs). 1) Most importantly, I think that in the current form the proof of the main theoretical result (Theorem 1) is wrong. The problem is with the reverse direction proving that any permutation invariant function can be represented in the form of Theorem 1. Specifically, Lines 142-159 construct matrix M which aggregates information about the graph edges.

Machine understanding of complex images is a key goal of artificial intelligence. One challenge underlying this task is that visual scenes contain multiple inter-related objects, and that global context plays an important role in interpreting the scene. A natural modeling framework for capturing such effects is structured prediction, which optimizes over complex labels, while modeling within-label interactions. However, it is unclear what principles should guide the design of a structured prediction model that utilizes the power of deep learning components. Here we propose a design principle for such architectures that follows from a natural requirement of permutation invariance. We prove a necessary and sufficient characterization for architectures that follow this invariance, and discuss its implication on model design.