Learn Before
Case Study

Decide on the next steps for a computer vision team reviewing a misclassified dev set.

Case context: A machine learning team is building an image classifier. During a recent evaluation, the algorithm misclassified several dev set images. The team starts manually reviewing these mistakes. They notice a large fraction of errors involve blurry images, but they currently have no idea how to improve the algorithm's performance on blurry inputs. Some team members suggest ignoring blurry images completely and only focusing on errors they know how to solve.

Question: Based on the principles of error analysis in Machine Learning Yearning, what should the team do regarding the blurry images, and why?

Sample answer: The team should not ignore the blurry images. They should still include them in their error analysis and count their frequency. The goal of examining misclassified examples is to build intuition about the most promising areas to focus on, not just to log errors they already know how to fix. By analyzing these blurry images, asking how a human would classify them, and understanding their underlying causes, the team might be inspired to come up with new solutions. Furthermore, knowing the exact fraction of errors caused by blurriness will help them prioritize whether it's worth dedicating a new project to tackle that specific issue in the future.

Key points:

  • Do not restrict analysis only to errors with known solutions
  • Goal is to build intuition about promising areas
  • Can inspire new solutions by asking how a human would label them
  • Helps figure out how promising different directions are for prioritization

Rubric: The student should advise against ignoring the blurry images, stating that error analysis shouldn't be restricted to easily fixable errors. They should note that analyzing these errors builds intuition, can inspire new solutions, and helps prioritize future efforts.

0

1

Updated 2026-05-27

Contributors are:

Who are from:

References


Tags

Machine Learning

Deep Learning

Supervised Learning

Dive into Deep Learning @ D2L

Data Science

Machine Learning Strategy

Machine Learning Yearning @ DeepLearning.AI

Related