Diagnose a report claiming 85% training accuracy and 95% dev/test accuracy.
Case context: A team reports that its algorithm achieves 85% accuracy on training examples it has seen and 95% accuracy on dev/test examples it has not seen.
Question: Based only on the source principle, what should the team diagnose about this report, and what performance must be addressed first?
Sample answer: The team should diagnose the reported relationship as inconsistent with the source principle. Dev/test performance is usually worse than training-set performance, so an algorithm achieving only 85% on seen examples cannot be achieving 95% on unseen examples. The team should first address and improve training-set performance before expecting strong dev/test performance.
Key points:
- The reported performance relationship is inconsistent with the source.
- The algorithm scores lower on seen examples than it allegedly does on unseen examples.
- Dev/test performance is usually worse than training performance.
- Training-set performance should be addressed first.
Rubric: Full credit requires identifying the report as inconsistent, connecting training data with seen examples and dev/test data with unseen examples, and stating that training-set performance must be addressed first.
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
What must occur before strong dev/test performance can reasonably be expected?
Dev/test performance is usually worse than training-set performance.
An algorithm must perform well on the _____ set before strong dev/test performance can be expected.
Match each performance concept to its source-based meaning.
Order the reasoning used to evaluate a training-versus-dev/test performance claim.
Explain why training-set performance sets a prerequisite for dev/test performance.
Diagnose a report claiming 85% training accuracy and 95% dev/test accuracy.
Why should training results be checked before expecting strong dev/test results?
Which reported result conflicts with the source's training-first principle?
An algorithm performing poorly on training data can still be expected to perform well on dev/test data.