Definition
A learning system improves through a loop. Data goes into a model, the model makes predictions, a loss measures how wrong they are, and that feedback adjusts the model. A model is just a function with adjustable numbers, called parameters.
In simple words
Learning cricket: you bat (prediction), see where the ball went (result), the coach says what was wrong (feedback), and you adjust your stance (parameters). Do this ten thousand times.
Why it matters
This one loop sits behind nearly every AI system, from a spam filter to ChatGPT. Learn it once, and every later algorithm is a variation on it.
Key terms
- Learning cycle
- The repeating sequence of data, model, prediction and feedback through which a system improves.
- Feedback
- A signal showing whether a prediction was right, used to improve the model.
- Feedback loop
- A situation where a system's outputs change the data it later learns from.
- Drift
- The gradual change of the real world away from the data a model was trained on.
- Monitoring
- Regularly checking a model in use to see whether it still performs well.
Explained in detail
The whole process in one picture
The topics of this module fit together as a cycle. Data is collected and prepared. A model is trained on it. The model makes predictions on new cases. Those predictions are compared with what actually happened, and that comparison is feedback. The feedback becomes new data, the model is trained again, and the cycle turns once more. Each stage depends on the one before, so a weakness anywhere travels all the way round.
The cycle in a product you know
Take a keyboard that suggests the next word. Its model was trained on a very large collection of text. As you type, it predicts three likely words. If you tap one, that is a sign the prediction was good. If you ignore them and type something else, that is a sign it was not. Signals like these, gathered from many users, are used to retrain the model, and the suggestions slowly improve. The same cycle runs in a music app (data: what you played; prediction: the next song; feedback: whether you listened or skipped) and in a map app (data: past journeys; prediction: arrival time; feedback: when you actually arrived).
When the loop feeds itself
The cycle has a hidden danger. A system's predictions can change the very data it will learn from next. Suppose a video app shows you cricket clips because you watched one. You watch more cricket, because that is what is on your screen. The app records that you love cricket and shows even more. It never finds out that you would also have enjoyed cooking videos, because it never offered any. The system's own choices have shaped the evidence, and its belief grows stronger without becoming more true. This is called a feedback loop.
Feedback loops can cause real harm in serious settings. If a system sends more inspectors to neighbourhoods where more problems were recorded in the past, the inspectors will record more problems there, simply because that is where they are looking. The data then appears to confirm the system's choice. Breaking such loops takes deliberate effort: showing people some things outside the prediction, and checking the system against information it did not create itself.
The world moves on
A model is trained on the past, and the present keeps changing. New slang appears, prices rise, fraudsters invent new tricks, a new road opens. When the pandemic began in 2020, shopping and travel patterns changed almost overnight, and many models that predicted demand suddenly became wrong. This gradual or sudden mismatch is called drift. It is the reason a model cannot simply be built once and left alone. Teams monitor their models in use, watch for falling performance, and retrain them with fresh data.
How it works
- Data: examples with right answers.
- Model: makes a prediction with its current parameters.
- Loss: measures the error.
- Update: changes the parameters to reduce the error.
- Repeat, thousands or millions of times.
Try it yourself
Run the learning cycle
A music app learns what you like. Follow it around the loop, and notice that it never really ends.
Run the learning cycle step by step
Put the steps of the AI learning cycle in order, starting with the data.
- 1Add the new example and its correct answer to the data
- 2Train the model on the examples
- 3Compare the prediction with what really happened (feedback)
- 4Train again, so the model improves
- 5Collect labelled examples (data)
- 6The model makes a prediction for a new case
Common mistakes
- Thinking the model is told the rule. It only ever gets feedback.
- Blaming the algorithm when the real problem is bad data.
Check yourself
Summary
- Data, model, loss, update, repeat.
- A model is a function with adjustable numbers.
- Feedback drives all the learning.
Finished reading?