Google

Showing training error and development error on the same graph makes it easier to estimate how development error may change as the training set grows. Looking at the full curve also reveals patterns such as high bias, high variance, or both more clearly than checking only the final point at the largest training-set size.

Why plot training and development error together on a learning curve?

A common sign of high bias is that, even with the largest available training set, the training error is still much higher than the target performance level. If the training error and development error are close together, the model is not showing much variance; the main problem is that it is underfitting rather than spreading its mistakes between train and dev.

Reading a High-Bias Learning Curve When the Train-Dev Gap Is Small

When training error is already near the best possible level, there is little opportunity to reduce bias. If dev error remains much worse than training error, the main room for improvement is reducing variance.

High Variance Clue from Learning Curves

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

True or False: Looking only at the error from the largest training sample gives the full story of how performance changes as data increases.

Comparing the training curve with the _____ curve can help you predict how the model will behave on unseen data.

Generalization Curves

Reading a Learning Curve

How to build and read a learning curve for a model

Question:
Explain why a model should be evaluated with an entire learning curve, showing training and dev error at several training-set sizes, rather than only at the final point using all available data.

Sample answer:
A single endpoint only tells you how the system behaves with the largest dataset you currently have. It does not show whether the model improves steadily, flattens out, or still has room to benefit from more examples. A full learning curve shows how both training and dev error change as more data is added. That broader view helps you judge whether more labeled data is likely to help and lets you estimate the dev-error trend with more confidence.

Key points:
- One endpoint summarizes only the largest available training set.
- The full curve shows behavior across multiple dataset sizes.
- Training and dev error together reveal learning trends.
- The dev curve can be projected more reliably when the whole curve is available.

Rubric:
A strong response should contrast the limited information from the final point with the richer trend information from the full curve, and it should explain that the curve makes extrapolation of dev error more reliable.

Why a full learning curve is more informative

Case context:
A team has trained a spam classifier on all 8,000 labeled emails currently available. Its training error and development error are known at that dataset size, but the team is considering spending several weeks labeling 4,000 additional emails.

Question:
What graph should the team create, and which error curves should it include so they can decide whether collecting more data is likely to help?

Sample answer:
The team should draw a learning curve that shows how both training error and development error change as the training set grows. For example, they could evaluate models trained on 1,000, 2,000, 4,000, 6,000, and 8,000 examples, then plot both curves on the same axes. Seeing the full pattern makes it easier to judge whether the development error is still dropping enough to justify labeling more data.

Key points:
- Create a learning curve over multiple training set sizes.
- Include both the training error curve and the development error curve.
- Use the development error trend to estimate whether more data is likely to help.

Rubric:
The answer must recommend plotting a learning curve across several dataset sizes instead of relying on a single final measurement. It must also name both training error and development error as the curves to include.

Choosing a plot to judge whether more labeled data is worth the effort.

Question:
What is the main advantage of placing the training error curve and the validation error curve on the same graph instead of inspecting them separately?

Sample answer:
It makes it easier and more confident to estimate the validation error trend.

Key points:
- A shared plot helps project the validation error more reliably.
- Comparing both curves directly increases confidence in that estimate.

Rubric:
The response should explicitly state that plotting them together improves confidence in estimating or projecting the validation error.

Learn Before

Related