equivariant-architecture-designer

To design an equivariant neural network respecting SO(3) or E(3) symmetry, provide a symmetry specification and task requirements to generate an architecture plan with appropriate layer patterns and non-linearities.

An equivariant neural network respects identified symmetry groups to ensure models remain data-efficient and robust to distribution shifts across images, point clouds, and graphs for tasks like classification or regression.

Implementation guidance for equivariant layers, non-linearities, and pooling in PyTorch is provided through library recommendations like e3nn and escnn, alongside evaluation criteria to verify equivariance.

Yes, architecture selection supports point clouds and graphs across tasks such as classification, segmentation, and regression by applying reference patterns and templates for symmetry groups like Cn and E(3).

The best way to verify equivariance is by applying the provided evaluation criteria to test that the identified symmetry groups are preserved through the network's equivariant layers and pooling operations.

Standard networks fail on distribution shifts because they lack symmetry constraints; designing architectures that respect identified symmetry groups like E(3) ensures models remain robust and data-efficient.