week8_summary.md

Week 8 Summary: Generalization, robustness, and process windows

Cross-Book Summary

1. The Generalization Problem

• Over/Underfitting: Balancing high variance (noise fitting) vs. high bias (too simple).
• Tradeoff: Minimizing total error = Bias² + Variance + Noise.
• Regularization: L1 (Lasso) and L2 (Ridge) penalize complex models.

2. Robust Validation

• Cross-Validation: K-Fold provides stable performance estimates.
• Leave-One-Out (LOO): For extremely small datasets.
• Stratification: Maintains representative class/property distributions.

3. Process Robustness & Windows

• Sensitivity Analysis: Quantifying output change vs. input perturbation.
• Process Windows: Finding optimal, stable regions insensitive to industrial noise.
• Robust Design: Choosing stability over a sharp performance peak.

90-Minute Lecture Strategy

Part 1: Reliability

• High accuracy is insufficient for deployment.
• Defining "Model Trust".

Part 2: Bias/Variance

• U-shaped error curve.
• Visualizing overfitting.
• Double Descent.

Part 3: Robust Validation Workflows

• K-Fold, Stratified, Grouped CV.
• Nested CV.
• Reliability metrics.

Part 4: Regularization

• L2 (Ridge) and L1 (Lasso).
• Early Stopping.

Part 5: Process Windows

• Mapping safe processing zones.
• Sensitivity analysis applications.
• Casting/3D Printing stability.

Quarto Website Update (Summary)

Summary for ML-PC Week 8:

• Shifts focus from performance to Model Reliability.
• Explores Bias-Variance tradeoff and generalization.
• Details robust validation (K-Fold, Stratified CV).
• Uses sensitivity analysis to map stable Process Windows.