How machine learning finds patterns — Computing, 14–17 years
In machine learning, a system adjusts its internal numbers from examples instead of receiving every rule directly. It can recognise patterns, but its answers depend on its data and do not equal human understanding.
Learning means adjusting from examples
A machine-learning model is a set of adjustable numbers that turns input into an output. During training, it sees examples and changes those numbers so its answers become closer to the examples’ known answers. After training, it applies the learned pattern to new input; it does not store a little person who understands the subject.
Why learn from data?
Some tasks are easy to describe in words, but others have too many messy cases for people to write every rule. Recognising a face or judging whether a sentence sounds like spam can involve subtle combinations. Machine learning was developed to find useful regularities in examples, especially when a clear hand-written rule is difficult to make.
Sorting messages by examples
Suppose a model receives 1,000 messages labelled spam or not spam. It notices that words such as “prize” and “urgent” often appear in spam, but it also learns from their combinations and the rest of each message. On a new message, it calculates a spam score of 0.92 and labels it spam. That score is a prediction, not certainty.
A pattern is not a cause
A reasonable mistake is to treat a strong correlation as an explanation. If training data show that umbrellas often appear with wet streets, a model can use umbrellas to predict rain, but umbrellas do not cause rain. A model may also rely on a misleading clue, especially when the examples are incomplete or biased. Prediction and understanding are different.
Useful predictions, careful decisions
Machine learning helps filter spam, suggest songs, translate text and spot unusual activity in medical images or factories. In each case, its performance depends on the examples and on what error matters most. A recommendation can be merely annoying, while a medical alert needs human checking and carefully chosen data.
Keep exploring
Other languages
Loading MyLeoNes™…