Two Kinds of AI Systems
Over the past three topics, we have looked at AI systems that reason through problems using explicitly defined rules. A production system applies productions — rules that a human wrote — to move from one state to another. A hand-crafted decision tree asks questions that a domain expert decided were relevant. A rule-based spam filter checks for keywords that a programmer identified as suspicious.
These systems work. For well-understood, relatively stable problems, explicit rules are often the right approach. But they have a fundamental limitation: someone has to write the rules. And for many of the most interesting and important problems, the rules are not known in advance, cannot be written down precisely, or change faster than any human could update them.
How do you write rules for recognizing a face? There is no rule that says "if pixels 412, 871, and 2,203 have these exact values, it is a face." Faces vary infinitely in lighting, angle, age, and expression. How do you write rules for translating a sentence from English to Mandarin? Natural language is too ambiguous, too context-dependent, and too vast for any explicit rule set to capture adequately.
Machine learning takes a different approach entirely: instead of writing rules, you show the system examples and let it figure out the rules itself.
What Learning Actually Is
In the context of AI, learning means this: a system adjusts its internal structure or parameters based on experience, such that its performance on a task improves over time without being explicitly reprogrammed for each improvement.
That definition has three important parts worth unpacking.
"Adjusts its internal structure or parameters." A learning system is not static. Something inside it changes as it encounters data. For a decision tree, that something is the tree structure itself — which attributes to split on and in what order. For a neural network (which we will study in Topic 6e), it is the numerical weights on thousands or millions of connections. Whatever the mechanism, learning means the system is different after training than it was before.
"Based on experience." The changes come from data — examples, outcomes, feedback. Not from a programmer manually adjusting the system. The system, in a meaningful sense, teaches itself from what it encounters.
"Performance improves without being explicitly reprogrammed." This is the key distinction from a rule-based system. When you update a spam filter's keyword list by hand, the programmer learned and updated the system. When a spam filter updates itself based on messages users flag, the system learned. The locus of learning has shifted from the human to the machine.
What a Model Is
The output of a learning process is called a model. A model is a compact representation of patterns extracted from training data — something that can be applied to new, previously unseen inputs to produce a prediction or decision.
The decision tree from Topic 6d is a model. It was built from thousands of labeled loan applications, but the tree itself is much smaller than that data — a few dozen nodes rather than thousands of rows. The tree has generalized from the examples: it has extracted the structure that predicts outcomes and discarded the rest. When a new loan application arrives, the model applies that extracted structure to produce a classification.
This is what separates a model from a database. A database stores all the training examples and answers a query by looking them up. A model distills the training examples into a structure that can answer questions about inputs it has never seen. The model can generalize. The database can only retrieve.
A model is what you get when a learning algorithm has finished its work. It is not the data, and it is not the algorithm — it is the pattern the algorithm extracted from the data. Deploying an AI system means deploying a model.
What Learning Is Not
Two things are sometimes confused with learning that are worth distinguishing.
Memorization is not learning. A system that stores all its training examples and matches new inputs against them exactly is not learning — it is retrieving. This is the overfitting problem from Topic 6d taken to its extreme. A model that has simply memorized its training data has not extracted any generalizable pattern. It will fail on any input that differs even slightly from what it has seen before. True learning produces a model that performs well on new, unseen data — not just on the examples it was trained on.
Lookup is not learning. Early chess programs beat human players by storing millions of board positions and their optimal responses. These systems were not learning — they were applying an enormous precomputed lookup table. Modern chess programs, by contrast, use reinforcement learning to develop strategies that generalize across positions, including positions they have never seen. The difference is not performance (both can beat humans) but whether the system has developed a generalizable model or is simply retrieving stored answers.
Three Fundamentally Different Approaches
Machine learning is not one thing. It is a family of approaches, and the differences between them are not matters of degree — they are matters of kind. The three major categories are organized around a single question: what information is available to the learning system as it trains?
- Supervised learning: The system trains on labeled examples — inputs where the correct answer is already known. The label supervises the learning process by telling the system when it is right and when it is wrong.
- Unsupervised learning: The system trains on unlabeled data. No one has told it what the "right" answer is. It looks for structure, patterns, and groupings in the data on its own.
- Reinforcement learning: The system learns through interaction with an environment. It tries actions, observes the consequences, and adjusts its behavior based on rewards and penalties. There are no labeled examples — just feedback from the environment over time.
Each of these three approaches is appropriate for different kinds of problems, and each has its own strengths, limits, and practical challenges. Reading 2 covers supervised and unsupervised learning in depth. Reading 3 covers reinforcement learning and then steps back to show how all three fit together as a conceptual map for the rest of Week 6 and Week 7.