Why More Training Data Does Not Fix a Large Training Error
Question: A speech classifier still has a training error much higher than the target error. Explain why collecting more training examples is not the right fix in this situation, and state what the engineer should concentrate on first.
Sample answer: When the training error is high, the main problem is usually high avoidable bias. Adding more training data is useful when the model has high variance, because extra examples can help it generalize better to new data. But if the model already performs poorly on the training set, more data will not make it better at fitting the examples it already has. The right next step is to improve the model's fit on the training set first, before expecting meaningful gains on validation or test performance.
Key points:
- More training data mainly helps with variance, not with high training error.
- A large training error relative to the target indicates high avoidable bias.
- Collecting more data does not substantially reduce bias.
- The engineer should first improve performance on the training set.
Rubric: Response should explain that adding more data is helpful for variance but not for a model that already has high training error. It must identify the problem as high avoidable bias and state that the engineer should focus on improving training-set performance before expecting better validation or test results.
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
An email-spam classifier has 10% training error and 11% development error, but the team wants to reach 3% error. What should they do first?
True or False: When a model has high training error because it does not fit the training set well, collecting more training examples is usually the main fix.
Effect of More Training Examples
Your model’s training error is 18%, and your goal is 6%. What should you address first?
More Data Quickly Solves Underfitting
More training data mainly helps with _____ issues, not with bias.
Match each learning situation to the most appropriate bias-or-variance description.
Order the steps for diagnosing and responding to a model that performs poorly on the training set.
A model has 18% training error, 19% development error, and the target error is 6%. What does this most strongly suggest?
True or False: If training error is still high, adding more data usually leads to big gains on the development and test sets.
If training error remains large, improve the _____ fit before expecting validation or test performance to rise.
Match each error pattern to the correct diagnosis and recommended action.
Prioritize the next improvement step when considering more training data.
Why More Training Data Does Not Fix a Large Training Error
Diagnosing a speech command classifier with high training and dev error
What Should Improve First Before Validation Results Are Expected to Move?