Overfitting
A model memorises the noise in its training data instead of the real pattern, so training error keeps falling while test or validation error rises. Lecture 4 calls this the high-variance failure mode: low training error, high test error.
When to use
Expect it whenever a model is too flexible for the data: a degree-10 polynomial on 30 points, an unrestricted tree grown until every leaf is pure, or a network trained for too many epochs. PS 02 (degree-5 polynomial features give a near-perfect training score but a hugely negative test score) and PS 07 test it. The lecture fixes are to regularise, get more data or simplify, and to stop when validation loss stops improving.
See DS Lec 4.