LLoCa vs L-GATr
Most previously published Lorentz-equivariant networks rely on specialized layers to achieve Lorentz-equivariance. Popular examples are
LorentzNet, a Lorentz-equivariant graph network using scalar and vector representations
PELICAN, a Lorentz-invariant graph network that uses general permutation-equivariant layers operating on Lorentz invariants
CGENN, a Lorentz-equivariant graph network using geometric algebra representations
L-GATr, a Lorentz-equivariant transformer using geometric algebra representations
Canonicalization, or specifically Lorentz local canonicalization (LLoCa), is an alternative approach that extends an existing backbone architecture with a Lorentz-equivariant canonicalization procedure to achieve Lorentz-equivariance. This section discusses the benefits of both approaches and helps to decide which one to use for your specific application. We focus on L-GATr as a representative of the specialized-layer approach because we compared it carefully with LLoCa-Transformers, but most arguments apply to other specialized-layer networks as well.
Disclaimer: Lorentz-equivariant networks are great, but when starting on a task from scratch we recommend to start with a non-equivariant baseline network first, for instance a vanilla transformer, and move to a Lorentz-equivariant network later on for the extra performance and robustness.
Benefits of L-GATr / specialized layers
Simpler, e.g. LGATrSlim fits into a single file. Canonicalization modifies the backbone architecture only slightly, but the Frames-Net and subsequent orthonormalization requires significant extra code.
Training dynamics are typically easier, because the interaction between the Frames-Net and backbone in LLoCa can lead to more complex training dynamics. For instance, we found in some cases that the Frames-Net in LLoCa overfits before the backbone does, even though the Frames-Net has far fewer parameters.
Benefits of LLoCa / canonicalization
Make any backbone architecture Lorentz-equivariant: We constructed LLoCa-Transformer, LLoCa-GNN, LLoCa-ParticleNet and LLoCa-ParT networks in our publications. LLoCa is particularly useful when your non-equivariant architecture already includes many tricks that would be time-consuming to re-implement in a Lorentz-equivariance-by-layers architecture like L-GATr, such as dynamic graph convolutions, attention mechanisms, or U-Net layouts. See Make any network Lorentz-equivariant for instructions and code examples for how to extend your favorite architecture with LLoCa.
Higher-order representations are not only straight-forward to include in LLoCa but also easily mixed in equal or unequal fractions. LLoCa builds on generic tensor representations of the Lorentz group through the
TensorRepsclass, e.g.16x0n+8x1n+2x2n+1x3nfor a direct product of 16 scalar, 8 vector, 2 second-rank tensor, and 1 third-rank tensor representations. Arbitrary higher-order representations are implemented, but using them comes at the cost of slower inference because of the additional matrix multiplications.Reduced resource constraints: For similar network size, we find that LLoCa networks typically require less GPU memory and FLOPs than comparable specialized-layer networks. This is because the backbone processes Lorentz-invariant features only, which are typically lower-dimensional and cheaper to process. The overhead from the Frames-Net and frame-to-frame conversions in message passing increases inference time by a factor of 1.5-2x in our current implementation, but based on our FLOPs analyses we believe that there is significant room for optimization that we want to explore in the future.
Flexible for studies on subgroup equivariance and data augmentation: Subgroup-equivariant networks such as SO(3)-equivariant or SO(2)-equivariant networks as well as data augmentation can be implemented easily as modifications in the Frames-Net. This allows for systematic studies of the impact of symmetry-aware designs, which would require entire new network designs when using the specialized-layer approach.