Examples
inputs + (maybe) correct answers
Learning algorithm
adjusts internal parameters
Trained model
makes predictions on new inputs
Machine learning is the practice of building systems that improve at a task by being shown examples, rather than by being explicitly programmed for that task. Instead of writing 'if pixel pattern X, output cat', you show the system thousands of cat photos and let an algorithm figure out the pattern itself.
The three main flavors
Supervised learning
- Given labeled examples (input, correct answer)
- Learn to predict the answer for new inputs
- e.g. email → spam/not-spam
Unsupervised learning
- Given unlabeled examples only
- Learn structure: clusters, compressed representations
- e.g. group customers by purchase behavior
A third flavor, reinforcement learning, learns by trial and error: an agent takes actions in an environment and receives rewards, gradually learning which actions lead to better outcomes — this is how game-playing AI and part of chatbot fine-tuning work.
Why this approach works so well
Some tasks are easy to describe with rules (compute a tax bracket) but impossible to describe that way (recognize a face, understand sarcasm). Machine learning shines exactly where the 'rule' is too complex, too fuzzy, or too unknown for a human to write down — but where we can gather many examples of correct behavior.
Deep learning, the subject of the next star, is simply the most successful family of algorithms currently used to do this kind of learning at large scale.