Machine Learning Basics: From Zero to Understanding

Machine learning is the part of AI where a system improves at a task by being shown examples, rather than by following rules a person wrote. There are three main paradigms, and understanding which one applies to a given problem tells you a lot about how that system actually behaves.

Supervised learning

You show the system many examples, each already labeled with the correct answer, and it learns to predict that label for new, unseen examples.

Example: show a model thousands of emails labeled "spam" or "not spam." It learns which patterns correlate with each label, so it can classify a brand-new email it has never seen.

This is the most common and best-understood form of machine learning. It requires labeled data, which is often the expensive part - someone (or something) has to produce those labels first.

Unsupervised learning

You give the system data with no labels and ask it to find structure on its own - grouping similar items together, or finding which factors actually vary independently.

Example: given a retailer's full customer purchase history with no predefined categories, a clustering algorithm might group customers into segments based on buying patterns, without anyone telling it what those segments should be in advance.

There's no "correct answer" to check against, so evaluating unsupervised learning is inherently more subjective than supervised learning.

Reinforcement learning

The system takes actions in an environment and receives a reward or penalty, then adjusts its behavior to get more reward over time - learning through trial and error rather than from a labeled dataset.

Example: a system playing a game receives points for winning moves and none for losing ones, and gradually learns a strategy purely from that feedback signal, without ever being told the "correct" move for a given position.

Reinforcement learning is also used to fine-tune large language models on human preferences after their initial training - a step often referred to as RLHF (reinforcement learning from human feedback), where the "reward" comes from human raters comparing model outputs rather than a game score.

How these connect to the AI you actually use

Modern large language models are trained in stages that draw on more than one of these paradigms: an initial stage that resembles unsupervised learning (predicting the next word across huge amounts of unlabeled text), followed by supervised fine-tuning on curated examples, followed in many cases by a reinforcement-learning stage to align the model's behavior with human preferences. See How Large Language Models Work for the full training pipeline.

A common beginner confusion, cleared up

"AI," "machine learning," and "deep learning" are often used interchangeably, but they nest inside each other: AI is the broadest goal (getting machines to do tasks that need intelligence), machine learning is one approach to that goal (learning from data instead of hand-written rules), and deep learning is a specific technique within machine learning that uses many-layered neural networks - the technique behind essentially every headline AI system today, including image recognition and large language models.