Definition
Plain language
Making extra training examples by tweaking or duplicating existing data.
As stated in the literature
Techniques that expand a dataset by transforming existing samples (cropping, noising, paraphrasing) to improve generalization; contrasted in this corpus with decomposing one trajectory into faithful per-step training samples, which reproduces real decision points rather than distorting data.
Why it matters: It stretches limited data so a model generalizes better, though the tweaks must stay realistic or they can distort what the model learns.
For example, flipping, cropping, and slightly recoloring a single photo of a cat turns one labeled image into dozens of training examples.
Heard on the show
“Wait — so that's data augmentation?”Episode 156 — Why More Human Demonstrations Made a Computer-Use Agent Worse