How should a team organize a large validation set with many mistakes?
Case context: A project team has a validation set of 8,000 examples. Its model gets 15% of them wrong, so there are about 1,200 errors to study. Reviewing every mistaken case by hand would be slow.
Question: What split of the validation set is most useful, and what should the team do with each part?
Sample answer: The team should divide the validation set into two clearly separated parts: one part for close human review and one part that is never used for that review. The reviewed part should be used to inspect failure patterns and identify common error sources. The untouched part should be reserved for choosing model settings and tracking performance. If the team sees that improvements on the reviewed part stop translating well to the untouched part, that is a sign it may be fitting too closely to the reviewed slice.
Key points:
- Make one explicit split between a reviewed subset and a reserved subset.
- Use the reviewed subset for hands-on error inspection.
- Keep the reserved subset free of that manual review.
- Use the reserved subset for model selection and parameter tuning.
- Compare the two subsets to spot overfitting to the reviewed examples.
Rubric: The response should recommend a clear two-part split, assign manual error inspection to one part, assign model selection or tuning to the other part, and explain that the separation helps reveal overfitting to the reviewed subset.
0
1
Tags
Machine Learning
Deep Learning
Supervised Learning
Dive into Deep Learning @ D2L
Data Science
Machine Learning Strategy
Machine Learning Yearning @ DeepLearning.AI
Related
Human-Review Dev Set
Blackbox Dev Set
Use the Full Dev Set When It Is Too Small to Split
Why split a development set into a review subset and a tuning subset?
Does examining part of a dev set more closely increase the risk of overfitting to it?
The hands-off portion of the dev set can still be used to tune _____.
Match each dev-set idea with its role or consequence.
Order the reasoning for managing a dev set that is too large to inspect fully by hand.
Explain how separating a reviewed subset from an untouched subset can reveal overfitting during error analysis.
How should a team organize a large validation set with many mistakes?
How are the two parts of a split dev set used?
Which result suggests a model has been tuned too closely to the hand-checked subset?
If a development set is too small to divide into separate analysis and tuning subsets, using the full set for both purposes is a reasonable choice.