Case Study

Prioritize the pipeline stage with the largest traced error share.

Case context: A team is building a pipeline that classifies news articles into topic labels. After reviewing 120 dev set mistakes at the component level, they trace 60 errors to the tokenizer, 30 to the encoder, and 30 to the final classifier.

Question: Which component should the team improve first, and why does the error-fraction estimate support that choice?

Sample answer: The tokenizer should be improved first. Component-level review assigns each mistake to one stage, which lets the team estimate how much of the total error each stage contributes. Because the tokenizer accounts for the largest traced share of errors, 60 out of 120, it is the best first target for reducing overall dev set error.

Key points:

  • Improve the tokenizer first
  • It explains the largest traced error share: 60/120 = 50%
  • Compare estimated error fractions to decide where to focus effort

Rubric: Identify the tokenizer as the first priority. Explain that it has the largest traced error fraction. Connect the decision to comparing error fractions across components.

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