The document discusses deep learning on point sets. Point sets have properties of being unordered, invariant to transformations, and having interactions among points. The PointNet architecture uses three key techniques: 1) a max pooling layer as a symmetric function to aggregate point information, 2) a structure to combine local and global information, and 3) two joint alignment networks to align input points and point features. The architecture provides permutation invariance and is robust to data corruption.