May we set cookies? Ours tells a returning reader from a new one; Google's and TikTok's tell them whether an advertisement brought you here and what it led to. Say no and the site is exactly the same, and nothing that identifies your browser is kept or shared. What we measure

Learn machine learning by overfitting on purpose

Start a machine learning plan, first lesson free

Machine learning · fitting2 HELD-OUT MISSES

Model complexity: 4 of 10

training2/14 wrongheld out2/14 wrong

Training error 2/14, held-out error 2/14. When those two track each other the model is describing the pattern rather than the sample. Push the complexity up and they will part company, which is the moment to stop.

Place points, then raise the complexity: training error falls to zero while held-out error climbs.

Everyone can recite that a flexible model fits training data better and generalises worse. Fewer can tell when it is happening to them, because the two errors usually get described, not watched. The block below shows both and lets you cause the problem on purpose, so you know it when you see it.

What people come here to do

You describe the goal in your own words and the plan asks questions back until it is one you would follow. Goals usually look like these.

“Move from calling libraries to understanding them”
“Get through a machine learning module with the maths I have”
“Prepare for interviews that ask why, not which”
“Understand the models my team already ships”

What a machine learning plan covers

There is no fixed syllabus. The plan is built around your goal and drops what it does not need. These are the topics it draws from.

Fitting, generalisation and the bias-variance trade-off
Regression and classification from first principles
Regularisation and cross-validation
Trees, ensembles and where they beat everything
Neural networks and what backpropagation is doing
Evaluation, leakage and honest baselines

Other subjects

Each one has its own lesson block to try.

See how it works and what it costs.