Supervised Learning

In simple terms

In supervised learning, you show a model a bunch of examples paired with the correct answer (“this picture is a cat”, “this email is spam”) and it learns to predict the answer on new examples it hasn’t seen. The word “supervised” refers to the labels — a teacher told the model what each example was.

More detail

The training loop:

1. Collect a dataset of input/output pairs.
2. Choose a model architecture (linear, tree, neural net, …).
3. Define a loss function measuring how wrong the model is.
4. Iteratively adjust the model’s parameters to reduce the loss — usually with gradient descent.
5. Evaluate on a held-out test set the model never saw during training.

Two main flavours:

• Classification — predict one of a fixed set of categories. Cat vs. dog. Spam vs. not. Digit 0–9.
• Regression — predict a real number. Tomorrow’s temperature. The selling price of a house.

Common pitfalls:

• Overfitting — the model memorises the training set and performs poorly on new data.
• Data leakage — information from the test set sneaks into training (e.g. preprocessing on the combined set).
• Imbalanced classes — accuracy is misleading when 99% of examples are one class.
• Distribution shift — production data differs from training data; performance silently degrades.

Modern industrial ML is overwhelmingly supervised: search ranking, recommendations, ad targeting, image moderation, machine translation, code completion. Even unsupervised techniques (like training a language model on raw text) are usually followed by supervised fine-tuning.

Why it matters

Most “AI that works” is supervised learning at scale: huge labelled datasets, well-understood architectures, careful evaluation. The gap between an interesting prototype and a useful product is almost always a labels and evaluation problem.

Real-world examples

• An email spam filter is a binary classifier trained on millions of labelled emails.

• A medical imaging model learns “tumour vs. healthy tissue” from radiologist annotations.

• A coding assistant is fine-tuned on accepted vs. rejected completions.

• Labeling is now an industry: companies like Scale AI and Surge employ tens of thousands of human raters to produce the labels for everything from autonomous driving to LLM fine-tuning.

Common misconceptions

• “Supervised learning needs perfect labels.” It tolerates noisy labels surprisingly well, especially at scale. It is much less tolerant of biased labels.
• “Bigger model is the answer.” Often more or better-labelled data is. Architecture matters less than people assume.

Learn next

How a trained model is actually used at run time: training and inference. The dominant model family for supervised learning: neural networks.