Loss Function
A function that measures how wrong a prediction is; training aims to make it as small as possible. Also called a cost or error function.
When to use
It is step 1 of training: for classification compare the predicted probability with the true label, for regression use the squared difference (familiar from RMSE). The lecture gives intuition only, with no loss formulas.
See Lec 8.