Learning rules from data
Machine learning is the practice of having a computer find patterns and rules automatically in large amounts of data, instead of a person programming each rule. The machine learns from the data — hence the name — and almost all current AI is built this way.
What learning actually means
Take spam detection. Writing rules by hand — if it contains this word, it is spam — is laborious and quickly circumvented. Instead you show the system a large volume of spam and legitimate mail and let it find the difference itself. That is training.
A trained model can then judge mail it has never seen with high accuracy. More data generally helps.
Three broad approaches
Supervised learning trains on labeled examples. Unsupervised learning finds structure such as groupings in unlabeled data. Reinforcement learning improves through trial and reward — the approach behind the Go programs that beat the world’s best players.
A trained model, as the phrase appears in the news, simply means one that has been through this and is ready to use.
Because training needs both large volumes of data and heavy compute, how much high-quality data you can assemble has become a decisive factor in how capable a model gets.
Where it already runs
- Recommendations in video and music apps
- Face unlock and automatic photo sorting
- Improvements in weather and traffic forecasting
- Credit card fraud detection
How the terms nest
AI is the broad category; machine learning sits inside it; deep learning is one method within machine learning. Advances in deep learning produced generative AI such as ChatGPT. It is the foundation everything modern rests on.