Case Study

Sizing a Manual Review Set for a High-Accuracy Defect Detector

Case context: A factory inspection team has improved its defect detector so that the overall error rate is now 1.5%. The lead engineer wants a manual review set with about 60 misclassified product images so the team can study fresh failure patterns.

Question: How many product images should you include in the manual review set to expect about 60 errors, and why?

Sample answer: You should include about 4,000 product images. Since the error rate is 1.5%, or 0.015, you divide the desired number of misclassified examples, 60, by 0.015. That gives 4,000. The calculation shows that when a system is already fairly accurate, you need a much larger review set to collect enough mistakes for analysis.

Key points:

  • Divide the target number of errors by the error rate: 60 / 0.015 = 4,000
  • Aim for roughly 60 misclassified examples to inspect manually
  • Explain that a low error rate requires a larger review set to gather enough mistakes

Rubric: Give full credit if the learner correctly computes 4,000 examples and explains that dividing the desired number of misclassified cases by the 1.5% error rate leads to a much larger dataset.

0

1

Updated 2026-08-12

Contributors are:

Who are from:

Tags

Machine Learning

Deep Learning

Machine Learning Strategy

Supervised Learning

Dive into Deep Learning @ D2L

Data Science

Machine Learning Yearning @ DeepLearning.AI