shortcut learning

**Shortcut Learning** is the **tendency of neural networks to learn simple, superficial features (shortcuts) that correlate with the target in training data but do not capture the true underlying concept** — the model finds an easy rule that works on the training set but fails on slightly different test data. **Shortcut Examples** - **Background Cues**: A cow classifier learns "green background = cow" because most cow images have grass backgrounds. - **Texture Over Shape**: CNNs prefer texture features over shape features — classify by texture patterns, not object shape. - **Spurious Correlations**: Hospital equipment in X-ray images correlates with diagnosis — model learns equipment, not pathology. - **Position Bias**: NLP models learn answer position rather than semantic content. **Why It Matters** - **Silent Failure**: Shortcut-trained models achieve high training/validation accuracy but fail silently on distribution shifts. - **Detection**: Requires careful out-of-distribution testing — standard accuracy metrics miss shortcuts. - **Semiconductor**: Models may learn tool-specific artifacts (chamber signature, recipe ID) instead of true process physics. **Shortcut Learning** is **learning the easy trick instead of the real skill** — models exploiting superficial correlations that don't generalize.