🧠Supervised Learning

Supervised learning is learning from examples that come with answers. Each training case is a pair: input, correct output. The model guesses, the loss compares the guess with the provided answer, and gradient descent nudges the parameters. Most deployed ML value in the world, fraud scores, medical triage, translation, speech-to-text, is supervised at its core.

Classification and regression

Two output shapes cover most of it. Classification picks a category: which digit, spam or not, which of 10,000 products. Regression predicts a number: tomorrow's demand, a house price, time-to-failure. The canonical classification exercise is MNIST: 70,000 handwritten digits, 28 by 28 pixels each, collected from US Census Bureau employees and high school students. It is the dataset this tutorial's 3D network trains on, and it has been the field's 'hello world' for three decades.

The sacred split

You never grade a model on the examples it trained on, for the same reason you never grade students on the exact questions they memorized. Data is split: a training set the model learns from, and a held-out test set it never sees until evaluation. MNIST's own test digits were even written by different people than its training digits, so passing it means reading genuinely unseen handwriting. When someone quotes a model's accuracy, your first question should always be: on what held-out data?