Learn AI · Level 2 · How AI works · step 3 of 16 · word 5 of 5
What is overfitting?
Overfitting is when a model learns its training examples too closely, like memorising answers, and then does badly on new data it hasn't seen.
In plain English
Think of a student who memorises last year’s exam answers. They ace that paper, then fail when the questions change. An overfitted model has learned the quirks of its training data instead of the real pattern.
An example
In a well-known 2016 study, researchers trained a model to tell huskies from wolves, using wolf photos that mostly had snow in them. It seemed to work, but it had really learned “snow means wolf”. Show it a husky in snow and it says wolf.
Watch out
To catch overfitting, teams test models on data kept aside during training. The same problem affects AI benchmarks: if test questions leak into the training data, a high score may just mean the model has seen the answers.
The data kept aside is called a test set: see train-test split.
Learn AI path · Level 2 · How AI works · step 3 of 16
Done? Next stepHow a machine learning model is built: from data to drift →See all of level 2