Machine Learning¶
working pillar 9
Classical ML and the craft of training models that generalize. An active growth area.
Pages¶
| Page | Covers |
|---|---|
| Classical machine learning | Bias-variance, regularised linear models, trees and boosting, SVM, clustering, PCA |
| Training craft | Optimizers, AdamW, schedules and warmup, regularisation, debugging a model that will not learn |
| Evaluation | Metrics under imbalance, leakage, split strategy, calibration |
| Tabular and time series | Feature engineering, forecasting, backtesting, baselines |
The theme¶
Across all four pages, the recurring failure is the same: an evaluation that is easier than reality. Preprocessing fitted before the split, a rolling feature that peeks one step ahead, accuracy quoted on 1% prevalence, a test set consulted a dozen times, a forecast never compared against persistence.
None of these are modelling errors. Every one of them produces a better number, which is why they survive review — and why the discipline in evaluation is worth more than any model choice.
Connections¶
- Deep Learning — when the data is perceptual rather than tabular.
- Scientific ML — ML where physics is available as structure.
- Foundations › Probability and statistics — the estimation theory underneath.