invariance testing

**Invariance Testing** is a **model validation technique that verifies whether the model's predictions remain unchanged under transformations that should not affect the output** — testing that the model has learned the correct invariances (e.g., rotation invariance for defect detection, unit invariance for process models). **Types of Invariance Tests** - **Geometric**: Rotate, flip, or shift defect images — prediction should be invariant. - **Unit Conversion**: Change units (nm to µm, °C to °F) — prediction should be identical. - **Irrelevant Features**: Change features that shouldn't matter (timestamp, operator ID) — prediction should not change. - **Semantic**: Paraphrase text inputs — NLP model prediction should remain stable. **Why It Matters** - **Robustness**: Models that fail invariance tests are fragile and may fail unexpectedly in production. - **Correctness**: If changing an irrelevant feature changes the prediction, the model has learned a spurious correlation. - **Systematic**: CheckList framework formalizes invariance testing as a standard model validation practice. **Invariance Testing** is **testing what shouldn't matter** — systematically verifying that the model ignores features and transformations it should be invariant to.