Case Study

Setting Holdout Sizes for a Very Large Image Dataset

Case context: A team is training an image classifier using 900 million labeled photos. A new engineer suggests reserving 30% of the data for evaluation because that is a common classroom rule of thumb.

Question: What is wrong with that proposal, and how should the team choose the fraction and absolute size of the dev/test sets?

Sample answer: That proposal is not appropriate for a dataset of this size. Keeping 30% of 900 million examples for evaluation would waste an enormous number of training examples. The team should use a much smaller share for the dev and test sets, far below 30%. The main rule is that these sets should be just large enough to give a dependable estimate of model performance; their size should be based on that need rather than on a fixed percentage.

Key points:

  • A 30% evaluation split is far too large for 900 million examples
  • The dev/test fractions should be much smaller than 30%
  • The right size is the minimum needed for reliable performance measurement

Rubric: Response should correctly identify that a 30% split is inappropriate for a very large dataset. It must state that the dev/test fraction should be much smaller than 30%. It must specify that the holdout size is chosen by the amount needed to evaluate performance reliably rather than by a fixed percentage.

0

1

Updated 2026-08-12

Contributors are:

Who are from:

Tags

Machine Learning

Deep Learning

Supervised Learning

Dive into Deep Learning @ D2L

Data Science

Machine Learning Strategy

Machine Learning Yearning @ DeepLearning.AI