Learn Before
Greatest Error Reduction From Fixing One Category
Question: A model makes 30 mistakes on a validation set. Nine of those mistakes come from one recurring issue: the input images are cropped too tightly. If a project eliminates that issue completely, what is the greatest fraction of the model's current errors that could disappear?
Sample answer: 30%
Key points:
- 9 of the 30 current errors belong to that category.
- 9/30 = 30%.
- Removing one category cannot eliminate more errors than belong to that category.
Rubric: The answer must state that at most 30% of the current errors can be removed.
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
After reviewing 150 misclassified drone-delivery images, package-label errors account for 9% of the mistakes, weather-haze errors account for 18%, and shadow-occlusion errors account for 22%. Which category should you likely prioritize?
If 120 validation errors are reviewed and 9 of them are bicycle mistakes, then fixing bicycle errors perfectly could remove at most 7.5% of the total errors.
Error analysis on 100 misclassified dev set examples shows that traffic-light mistakes can eliminate _____ of the errors at most.
Match each error group to the statement that best describes why it deserves attention.
Put the steps in order for using error-category counts to decide where to focus next.
What does an error category's share of dev-set mistakes tell you?
Should the easiest error category always get top priority in error analysis?
Error-category frequency gives an upper bound on improvement.
Match each error-analysis term with its meaning.
Put the analysis steps in the order that justifies not prioritizing one low-frequency error type.
Explain How Error Category Frequency Guides Project Prioritization
Choosing which error categories to fix first
Greatest Error Reduction From Fixing One Category