Short Answer

Why does noisy dev-set labeling matter more as a model improves?

Question: In a machine learning project, why can label mistakes in the dev set become a bigger problem after the model gets much better?

Sample answer: As model performance improves, the number of genuine mistakes in the dev set goes down. If the label noise stays the same, a larger share of the remaining measured errors comes from bad labels rather than from the model itself. That makes the evaluation less trustworthy, so cleaning up the dev-set labels becomes more valuable.

Key points:

  • Model errors decrease as performance improves.
  • Fixed label noise takes up a larger share of the remaining errors.
  • This weakens the reliability of the evaluation metric.
  • Improving label quality becomes worthwhile.

Rubric: The answer must explain that, as true model errors shrink, mislabeled dev examples account for a larger portion of the remaining error and make evaluation less reliable.

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