Add Week 3 slides link and expand summary

Link the unit 3 reveal.js deck on pelzlab.science from the Week 3 entry,
and broaden the summary to match what the deck actually covers
(transformation toolbox, bias-variance/regularization, error measures
for regression and segmentation) on top of the existing data integrity
and leakage focus.

Author: PhilippPelz

_freeze/site_libs/bootstrap/bootstrap-a63378bca81584060b6e09effc601ccc.min.css

@@ -204,12 +204,14 @@ Fourier inspection of micrographs; effects of sampling and filtering.
 
 *Lecture: Tuesday, 28.04.2026, 14:15-15:45 | Exercise: Thursday, 30.04.2026, 16:15-17:45*
 
+**Slides:** [Open](https://pelzlab.science/public_presentations/ml_for_characterization_and_processing/unit03_data_quality/01_intro.html)
+
 - Annotation uncertainty and inter-annotator variance.
 - Train/test leakage in materials workflows.
 - Why “good accuracy” often means a broken pipeline.
 
 **Summary:**
-This unit focuses on the most critical and often overlooked part of the ML pipeline: data integrity. We discuss systematic data cleaning and normalization techniques while highlighting the unique challenges of labeling experimental materials data, such as inter-annotator variance. A major focus is on **Data Leakage**, specifically how spatial and physical correlations in materials samples can lead to deceptively high model performance. We introduce robust validation strategies to ensure models generalize to truly unseen data.
+This unit covers the often-overlooked half of an ML pipeline: data integrity, validation, and how performance is measured. We start with the measurement chain and systematic **data cleaning** — handling missing values, outliers, and duplicates with a "fix at source" mindset. We then build the **transformation toolbox**: centering, min–max and z-score scaling, physics-aware non-dimensionalisation, log transforms, differentiation, and frequency-domain views (FFT, triggering for time series). On the supervision side we examine **labels and uncertainty** — inter-annotator variance, probabilistic labels, and a Bayesian view of priors, likelihoods, and posteriors — and then formalize the **bias–variance** tradeoff with parsimony and regularization. A major focus is **Data Leakage** in materials workflows (pre-processing, temporal, and group/spatial), tackled with proper holdout, K-fold, LOOCV, and stratified validation. We close with the **error measures** that decide what "good" actually means: MAE/MSE/RMSE and $R^2$ for regression, and confusion matrices, precision/recall, F1/Dice, IoU, and categorical cross-entropy for classification and segmentation.
 
 
 **Exercise:**