The actual mechanics, minus the maths fear
Training a model means starting with random guesses, measuring the error against known answers, and nudging internal numbers to reduce that error - millions of times. Nothing mystical: it is guided trial and error at enormous speed.
Suppose you want to predict which customers will buy bread on Fridays. Features are the inputs you measure: day, payday week, weather, past purchases. The label is the answer: bought or not. Supervised learning learns a mapping from features to labels using examples where the answer is already known.
Supervised learning learns from labelled examples. Unsupervised learning finds structure with no labels - like grouping customers who buy similar baskets. Reinforcement learning learns from reward and penalty, the way game-playing agents learn. And the reason we hold out a test set: a model must be judged on examples it never saw, otherwise it is memorising, not learning.
Iteratively adjusting a model to reduce error on known examples.
The measured inputs / the known answer being predicted.
Learning a mapping from labelled examples.
Memorising training data so well that new data fails.
Tools: Notebook
