Memory-Based Graph Networks

Data generated by Autodesk software such as Fusion and Revit models are complex and non-Euclidean data that can be represented as graphs. The benefits of such representation are that graphs can abstract complex data into intuitive graphical representation and also can carry other types of information such as semantic, visual, and geometric information in addition to the structure. Moreover, graphs are actionable and users can simply tweak them by adding/removing nodes or edges. This project aims to categorize graphs with high precision and in the meanwhile learn good representations of meaningful sub-structures. This can then be used to retrieve Revit models that have similar kitchen layout or fusion models that use a specific sub-assembly in their design which in turn can accelerate the design process by suggesting similar examples and potentially auto-complete the incomplete designs.

Abstract

Memory-Based Graph Networks

Amir Hosein Khasahmadi, Kaveh Hassani, Parsa Moradi, Leo Lee, Quaid Morris

International Conference on Learning Representations 2020

Graph neural networks (GNNs) are a class of deep models that operate on data with arbitrary topology represented as graphs. We introduce an efficient memory layer for GNNs that can jointly learn node representations and coarsen the graph. We also introduce two new networks based on this layer: memory-based GNN (MemGNN) and graph memory network (GMN) that can learn hierarchical graph representations. The experimental results shows that the proposed models achieve state-of-the-art results in eight out of nine graph classification and regression benchmarks. We also show that the learned representations could correspond to chemical features in the molecule data.