Machine learning is the most widely used approach in modern AI. Understanding what it is — and how it differs from traditional software — helps you make sense of how AI systems are built and why they behave the way they do.
What Machine Learning Is
In traditional software, a developer writes explicit rules: if this condition is true, do this action. The program follows those rules exactly.
In machine learning, a developer provides data and a learning algorithm. The system analyzes the data, identifies patterns, and builds a model — a mathematical representation of those patterns. That model can then make predictions or decisions on new data it has not seen before.
The key difference: the rules are not written by a person. They are discovered by the system from examples.
Models and Training Data
A model is the output of a machine learning process — a set of learned parameters that the system uses to make predictions. Training a model requires training data — a collection of examples the system learns from.
The quality and representativeness of training data directly affects model quality. A model trained on incomplete, biased, or inaccurate data will produce incomplete, biased, or inaccurate results. This is why data quality principles from Data 101 apply directly to AI.
Three Common Approaches
- Supervised learning — the training data includes both inputs and correct answers. The system learns to map inputs to outputs. Example: a model trained on thousands of labelled images learns to classify new images.
- Unsupervised learning — the training data has no labels. The system finds structure on its own. Example: grouping customers into segments based on purchasing behaviour without pre-defined categories.
- Reinforcement learning — the system learns by trial and error, receiving feedback on whether its actions lead to good or bad outcomes. Used in game-playing systems and robotics.
Training vs. Using a Trained Model
Training a model is computationally intensive and happens before deployment. Using a trained model — called inference — is much faster and is what happens when you interact with an AI tool.
When you use an AI tool, you are not training it. You are querying a model that was already trained on a large dataset, often by a separate organization.
Classification and Prediction
Two of the most common machine learning tasks are:
- Classification — assigning an input to a category. Is this email spam or not spam? Is this image a cat or a dog?
- Prediction — estimating a numerical value. What will this property sell for? How many units will we sell next month?
Limitations and Uncertainty
Machine learning models are probabilistic — they produce outputs based on statistical patterns, not certainties. A model can be highly accurate on average while still being wrong in specific cases. Models can also fail in unexpected ways when they encounter data that differs significantly from their training data.
A model is only as good as its training data. Errors, gaps, and biases in training data become errors, gaps, and biases in model output. Related learning: Data Cleaning 101 and Data Best Practices 101.
- Machine learning systems discover rules from data rather than following hand-written rules
- Training data quality directly determines model quality
- Supervised, unsupervised, and reinforcement learning are three common approaches
- Inference is using a trained model — not training it
- Models produce probabilistic outputs, not guaranteed correct answers