Machine Learning Training, from scratch

Train/val/test splits, loss curves, overfitting, regularisation, the things every ML course glosses over.

A from-scratch tour of how a model actually learns. We start from one feature and one weight, hand-derive the gradient, then build up to an honest train() loop.

What’s in here

• the train / validation / test split, and why a single split is dangerous
• writing a training loop in pure NumPy (no model.fit)
• reading a loss curve: when to stop, when to worry
• overfitting in three lines of code, and three regularisation knobs that fix it
• early stopping vs L2 vs dropout — when each one is the right tool

Prerequisites

• Python (you’ve written a for loop)
• Comfortable with NumPy arrays
• High-school calculus (we use one chain rule)