Machine Learning Fundamentals · free preview

Loss, Training Sets and Validation Sets

Training needs a numerical objective and a split of the data, because fitting well on what we fit is not the same as doing well on what we have not seen.

1. The need for a number

The need for a number

2. Two common losses

Two common losses

3. The split that makes the number honest

The split that makes the number honest

4. Why the training loss is not the answer

A model with enough parameters fits the training set perfectly, driving the training error to zero, and performs poorly on new examples from the same source. Which statement explains the gap between the training error and the error on new examples?

5. The error rate is hard to optimize

For classification, the most direct objective is the error rate, the fraction of examples the model gets wrong. State the property of the error rate that makes it hard to minimize directly with gradient methods: it is ____.

Practice this interactively

Free account · instant grading · spaced review that schedules itself.

Start this course for free