Diagnose whether a team should expect more data alone to reach its desired error level.
Case context: A team examines its learning curves and finds that training error is above its desired performance level. Its dev error is also higher than its training error. The team proposes collecting more training data as its only response.
Question: Using the source observations, diagnose whether the proposal is sufficient and justify your decision.
Sample answer: The proposal is not sufficient. As more training data is added, training error can only stay the same or get worse, so it cannot be expected to fall to the desired level. Because dev error is usually higher than training error—and is higher in this case—more data alone is also not expected to bring dev error to the desired level.
Key points:
- Training error is already above the desired level.
- Adding data cannot lower training error in the stated learning-curve pattern.
- Dev error is higher than training error.
- More data by itself is therefore insufficient.
Rubric: The answer should reject the proposal as sufficient and justify the diagnosis using both the behavior of training error and the relationship between dev and training error.
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
Why is adding more data alone insufficient when training error exceeds the desired performance level?
More data alone can bring dev error to the desired level even when training error remains above that level.
As training data increases, training error can only stay the same or _____.
Match each learning-curve element to its role in diagnosing why more data is insufficient.
Order the reasoning used to conclude that more data alone will not reach the desired performance.
Explain how the training- and dev-error curves prove that more data alone is insufficient.
Diagnose whether a team should expect more data alone to reach its desired error level.
What two observations support the conclusion that more data alone will not fix this learning curve?
Which learning-curve pattern most directly rules out more data as a sufficient solution by itself?
If training error stays unchanged as data is added, dev error can still be expected to cross below a lower desired level.