Google

A learning curve can show both high bias and high variance when the training error is far above the target level and the development error is much higher than the training error. In that situation, the model needs changes that address both underfitting and instability.

Interpreting Training and Dev Error When Both Fit and Stability Are Poor

Recognizing When Both Bias and Variance Are High

If a model has both high bias and high variance, its training error and development error are usually about the same.

Learning Curves That Point to _____

Reading Training and Validation Patterns

Diagnosing Bias and Variance in a Model

Question:
Explain what the training and development errors typically look like when a model has both high bias and high variance, and state the broad direction for improving the model.

Sample answer:
When a model has high bias, the training error stays high, so performance on the training set is still far from what we want. When it also has high variance, the development error is much worse than the training error, showing that the model does not generalize well. Seeing both patterns at once means the model needs changes that lower both underfitting and overfitting, not just one or the other.

Key points:
- Training error remains much worse than desired, which signals high bias
- Development error is much higher than training error, which signals high variance
- The model must be improved in ways that address both problems

Rubric:
A strong response must describe the high training error, the large gap between development and training error, and the need to reduce both bias and variance.

What a learning curve suggests when both underfitting and overfitting are present

Situation:
A team is checking a model that predicts weekly store demand. Their target error rate is 5%. On the training set, the model’s error has leveled off at 18%. On the development set, the error is 31%.

Question:
What do these results suggest about the model’s bias and variance, and what should the team try to do next?

Sample answer:
The model shows high bias because its training error is still far above the target level. It also shows high variance because the development error is much worse than the training error. The next step is to work on reducing both bias and variance.

Key points:
- High bias is indicated by 18% training error compared with a 5% target.
- High variance is indicated by 31% development error compared with 18% training error.
- The recommended direction is to reduce both sources of error.

Rubric:
The response must correctly identify high bias from the large training error, identify high variance from the training-versus-development gap, and recommend reducing both.

Evaluating a Sales Forecast Model

Question:
What does it suggest when a model's training error is high and its validation error is much higher?

Sample answer:
It suggests a variance problem in addition to the existing bias. The large gap means the model fits the training data better than the validation data.

Key points:
- Validation error is much larger than training error.
- That gap indicates variance.
- High training error means bias is also present.

Rubric:
The answer should explicitly state that the validation error is much higher than the training error and that this points to variance.

Learn Before

Related