Choosing a small error review sample.
Case context: A team has launched a text classifier and checked it on the dev set. The model produced 360 wrong predictions. One teammate proposes spending the next several days examining every mistaken example before deciding what to improve.
Question: If you are using an Eyeball dev set, what would you advise instead, and what is the reason?
Sample answer: I would start by examining a random set of about 50 errors rather than trying to inspect all 360 at once. A group of roughly 50 is usually enough to reveal the main patterns behind the mistakes. Looking through every error first would waste time, because the important failure types should become visible much sooner.
Key points:
- Suggest a random review of about 50 errors.
- Explain that this size is enough to uncover the main failure patterns.
- Note that reviewing all 360 immediately is not a good use of time.
Rubric: The answer should advise sampling about 50 errors and explain that this is enough to identify the main failure patterns while avoiding the time cost of checking all 360 examples.
0
1
Tags
Machine Learning
Deep Learning
Machine Learning Strategy
Supervised Learning
Dive into Deep Learning @ D2L
Data Science
Machine Learning Yearning @ DeepLearning.AI
Related
What does inspecting roughly 50 validation mistakes usually reveal?
A small sample of dev-set errors can reveal the main failure modes.
With about _____ manually reviewed mistakes, you can usually identify the main error patterns.
Match each error-analysis term to its description.
Order the steps for manually reviewing mistakes in a small validation sample.
Why Review a Small Set of Mistakes?
You must examine several hundred validation errors before you can tell what the main failure modes are.
Looking through about 50 manually sampled mistakes usually reveals the main error _____.
Match each manual error-review element to its role in diagnosing recurring mistakes.
Order the steps for deciding whether a manual review of dev-set errors is useful.
Why review about 50 errors in a manual error check?
Choosing a small error review sample.
How many validation mistakes should be checked to find the main error patterns?