Case Study

Prioritizing Error Categories in a Vision Classifier

Case context: A team has completed an error analysis on 100 misclassified dev set examples from their image classifier. They discover that "Dog" mistakes account for 8% of the errors, while "Great Cat" and "Blurry" image errors make up the vast majority of the remaining errors.

Question: Based on Andrew Ng's guidelines for error analysis, what should the team diagnose or decide regarding their next steps to improve the classifier, and why?

Sample answer: The team should decide to focus their efforts on addressing the "Great Cat" or "Blurry" image errors rather than the "Dog" mistakes. Since Dog errors only account for 8% of the misclassifications, a project to fix them would eliminate at most 8% of the total errors. Working on the other categories provides a much larger opportunity to eliminate more errors.

Key points:

  • Decide to focus on Great Cat or Blurry image errors.
  • Working on Dog errors can eliminate 8% of the errors at most.
  • Prioritizing higher-frequency categories helps eliminate more errors overall.

Rubric: The response must diagnose that the team should prioritize "Great Cat" or "Blurry" errors over "Dog" errors, justifying the decision by stating that fixing "Dog" errors has a strict upper limit of an 8% improvement.

0

1

Updated 2026-05-27

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