Data science, minus the course-seller fluff.

The version of the field that actually gets used: framing the problem, getting the data trustworthy, picking the simplest model that works, then explaining the result to people who don’t care about your architecture.

Error plotted against model complexity. The error on the training data falls all the way down. The error on unseen data falls, flattens out at a lowest point in the middle, then climbs again. as good as it gets unseen data training data model complexity error The figure comes from Overfitting and the bias-variance tradeoff →
error on data the model has never seen error on the training data Past the dot, a stronger model fits the training rows better and predicts worse. 32 of the 76 articles carry a drawing like this one, made for the one point it has to make.

Where each track begins

The full reading order groups all 33 Foundations articles by what each one assumes you already know.

The whole series, in reading order →

Latest articles

All articles →