Search Articles

Encoding Structure in Models #

• CNNs: Using spatial structure
  • The core idea is around convolution – this preserves spatial information
  • This works because we represent the data as a 2D array as the input to the model
• Graphs as a Structure for Representing Data
  • Graphs are a really powerful way to represent data
  • Many real-world data (such as networks) can’t be encoded in Euclidean data structures
• Graph Convolutional Networks (CGNs)
  • The idea is similar to standard CNNs – the convolution is slid over the different nodes + neighbors and extracts features from the local neighborhood of the graph
  • Graph encoding is a new technique that can be used to datasets that naturally have a graph structure
    • e.g. Molecular discovery

Applications of Graph Neural Networks #

• Drug discovery – discover novel antibiotics
• Traffic prediction – predict traffic patterns
  • This led to significant ETA improvements in Google Maps
• COVID-19 Forecasting
  • This was modeled as spatio-temporal data

Learning From 3D Data #

• Learning from point clouds!
  • Unordered sets of points in 3D space
• You can extend graph neural networks to point clouds
  • The idea is to use a graph to represent the point cloud
    • You can dynamically construct a mesh based on the point cloud