Choose the component to improve when most dev-set mistakes come from one stage.
Question: A review of 90 misclassified validation examples shows that 72 errors come from a lane-marking detector and 18 come from a traffic-signal classifier. Which component should you improve first?
Sample answer: Improve the lane-marking detector first, because it accounts for most of the errors.
Key points:
- The lane-marking detector is the priority.
- It is responsible for 72 of the 90 errors, far more than the traffic-signal classifier.
Rubric: Full credit for identifying the lane-marking detector as the component to prioritize because it produces the larger share of mistakes.
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
Example Attribution Can Support a Second Pass of Debugging
Which component should receive the most attention if 72 of 80 dev-set mistakes come from the signal preprocessor?
A single source of most validation errors should receive priority.
Identify the component with the most assigned errors
Interpreting Component Error Counts
Order the steps for using component error counts to prioritize pipeline fixes.
A speech-recognition pipeline shows 9 times as many dev-set mistakes in the wake-word detector as in the speaker-id module. What should the team do next?
If one subsystem is responsible for 10 out of 100 validation mistakes, it should be the first area the team improves.
Reviewing 100 incorrect validation predictions and assigning each one to a pipeline _____ shows which step should be improved first.
Match each pipeline error-analysis term to its correct description.
Order the steps used to decide which pipeline component deserves improvement first.
Use component error counts to set improvement priorities.
Using component error counts to decide where to improve a document-processing pipeline.
Choose the component to improve when most dev-set mistakes come from one stage.