symmetry-validation-suite

To test equivariance in machine learning models, apply predefined group transformations to inputs and measure output discrepancies. This suite executes empirical invariance and equivariance checks to generate quantitative validation reports.

Symmetry validation for neural networks is the empirical process of verifying if a model maintains invariance or equivariance under specific group operations. It ensures claimed symmetries hold across data shifts to guide architecture design decisions.

Verify rotation invariance by sampling rotation transformations from group operations and comparing model outputs before and after the shift. The suite provides test protocols and quantitative metrics to execute this empirical validation.

No specific framework is required to test permutation invariance, as the suite operates through protocol-driven testing without dependencies. It applies transformation sampling strategies to assess symmetry validity across different model architectures.

Empirically test symmetry hypotheses before adding architectural constraints when you need to confirm if claimed symmetries like E(3) or SO(3) equivariance actually hold. The suite generates diagnostic reports to guide hard constraints, soft constraints, or augmentation strategies.