Diagnose a field-to-test mismatch in apple inspection photos
Case context: A team trains a model to detect bruising on apples from photos. In production, the photos are taken by workers using handheld phones in dim storage rooms. However, the dev and test sets mostly contain bright, carefully staged lab photos.
Question: What is the main problem with the current evaluation setup, and what should the team do to correct it?
Sample answer: The dev/test sets do not reflect the distribution the model will face in real use. The team should rebuild the dev and test sets so they better match the actual production photos.
Key points:
- The production photos are different from the dev/test photos.
- The evaluation data is not representative of the real operating conditions.
- The dev and test sets should be updated to match the real data distribution more closely.
Rubric: The student must identify that the dev/test set is not representative of the real data and recommend updating the dev/test sets.
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
When the Development Set No Longer Matches Real User Data
If a dev/test set does not reflect the distribution the model will face after deployment, what should the team do?
True or False: A validation set drawn from a very different population than deployment data can still reliably show which model changes will help in the real world.
When the validation set does not match deployment conditions, the recommended action is to _____ the validation set.
Identify the distribution concepts used in model evaluation.
Arrange the actions a team should take when a validation set stops predicting real-world performance.
Why a Non-Representative Dev/Test Set Can Mislead Model Improvement
True or False: Evaluation data should match the conditions the final system will face in practice.
Core idea about dev and test data
Match each situation to the best action for dev/test data.
Order the reasoning steps showing why a development set that misses the real distribution should be revised.
Why mismatched evaluation data can mislead a project
Diagnose a field-to-test mismatch in apple inspection photos
What should a team do if its evaluation data no longer reflects real-world use?