How machines learn
Artificial intelligence is not a robot brain and it is not magic. At its core it is software that learns patterns from examples instead of being told exactly what to do. This lesson gives you every word you need for the rest of the course. By the end you will know what data, a feature, a label, and a model actually are, and you will have met them all through football.
Old software follows rules. AI learns them.
Think about a normal app, like a calculator. A programmer sat down and wrote every rule by hand: "when the user presses +, add the two numbers." The calculator never learns anything. It just follows the rules it was given, forever.
Machine learning flips that around. Instead of a human writing the rules, you show the computer thousands of past examples with the answers already filled in, and the computer works out the rules by itself. Nobody tells it "a bigger ranking gap means a higher chance of winning." It notices that pattern on its own, from the examples.
A human writes the rules. Rules + data go in, answers come out. It cannot handle anything the human did not plan for.
A human provides examples with answers. Data + answers go in, and the rules come out. The computer discovers the pattern itself.
Imagine learning to cook rice and peas without a recipe. Nobody hands you exact amounts. You just cook it many times, taste the result, and adjust. Too dry? More coconut milk next time. That is machine learning: many attempts, checking the result, adjusting. The "recipe" you end up with is the model.
Data is just a table of examples
Everything a model learns from is data, and data is almost always a table: rows and columns, like a spreadsheet. Each row is one example. Here, one row is one past football match.
| Ranking gap | Recent form | Home? | Did they win? |
|---|---|---|---|
| 26 | +3 | Yes | Win |
| 5 | −1 | No | No |
| −10 | +2 | Yes | Win |
| 30 | +5 | Yes | Win |
| −20 | −3 | No | No |
Look at the columns. The first three describe the match before it is played: how much stronger one team is (ranking gap), how well they have been playing lately (recent form), and whether they are at home. The last column is the thing we want to predict: did they actually win?
A feature is one input column, one fact about the example that might help predict the answer. Ranking gap, recent form, and home advantage are the three features in our table.
The label is the column you want to predict, the answer that is already filled in for past examples. Here the label is "Did they win?" You learn from matches where you already know the label, so the model can spot which features tend to go with a win.
Features are the clues. The label is the answer. A model studies old cases where it can see both the clues and the answer, so that later, when it only has the clues for a new match, it can guess the answer.
So what is a "model"?
A model is the finished machine that takes the features of a brand-new example and produces a prediction. You feed today's match facts in one end, and a prediction comes out the other.
The two models in this course, logistic regression and decision trees, are just two different designs for that middle box. They take the same features and reach a prediction in two very different ways. Learn both and you have basically learned this whole course.
Two kinds of prediction
There are two different questions you could ask about a match, and they need two different kinds of answer.
"How many goals will they score?" The answer is a number: 0, 1, 2, 3. Predicting a number is called regression.
"Will they win, yes or no?" The answer is one of a few choices. Predicting a category is called classification.
This course is mostly about the second kind: predicting a yes/no result. Logistic regression (Lesson 2) answers it with a probability, and decision trees (Lesson 3) answer it by asking questions. Both are classification models.
Probability: putting a number on "not sure"
Probability comes up a lot from here on, so let's be clear what it means. A probability is a number between 0 and 1 that says how likely something is. We usually write it as a percentage instead.
A probability of 0 means it will never happen. 1 (or 100%) means it definitely will. 0.5 means it is a 50/50 coin flip. Everything real about football lives somewhere in between.
Saying "Argentina has a 78% chance of winning" is far more useful than just "Argentina will win," because it is honest about doubt. A good model does not pretend to be certain. It gives you a probability, and you decide what to do with it.
How a model actually learns: guess, check, adjust
A model does not start out smart. It starts out random, making terrible predictions, and improves in a loop that repeats thousands of times. That loop is the part most people never realise is happening.
That is called training. Every example the model sees is a chance to check how wrong it was and shrink that error a little. Do it enough times and a random guesser turns into something genuinely useful. When you press "Fit" in the labs coming up, you are watching exactly this loop play out.
When you train a model on examples that already have the answer (the label) filled in, it is called supervised learning. The known answers "supervise" the learning by telling the model when it is right or wrong. Both models in this course learn this way.
- Data is a table of examples (rows) described by columns.
- Features are the input clues; the label is the answer you predict.
- A model takes features in and gives a prediction out.
- Classification predicts a category (win / no win); regression predicts a number.
- A probability (0 to 100%) is how a model expresses "how sure" it is.
- Training is the guess-check-adjust loop; using known answers makes it supervised learning.
Next you meet your first real model, and its famous S-shaped curve.