A team cannot confidently project its dev-error curve. What analysis should it add?
Case context: A machine-learning team has plotted dev error against the amount of training data. Team members find it difficult to predict exactly where the dev-error curve would go if they obtained more data.
Question: Diagnose the limitation in the team's analysis and decide which additional plot it should use.
Sample answer: The team is relying only on the dev-error curve, which can be difficult to extrapolate. It should also plot training error because that additional plot can help estimate the impact of adding more data.
Key points:
- The current analysis uses only the dev-error curve.
- The difficulty is extrapolating that curve to more data.
- The team should add a training-error plot.
- Training error can help estimate the effect of adding data.
Rubric: The response should diagnose reliance on dev error alone, state that exact extrapolation is difficult, recommend a training-error plot, and connect that plot to estimating the impact of more data.
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
Training Error Usually Increases with Training Set Size
Training Error Is Usually Lower Than Dev Error
Why should training error be plotted when evaluating whether more data may help?
A dev-error curve alone may be difficult to extrapolate to larger training sets.
The additional plot used to estimate the effect of more data is _____.
Match each learning-curve element to its role in estimating the effect of more data.
Order the reasoning process for assessing the possible impact of adding data.
Explain why training error strengthens an analysis of whether more data may help.
A team cannot confidently project its dev-error curve. What analysis should it add?
What limitation motivates adding a training-error plot?
Which analysis best follows the source when considering a larger dataset?
Training error is intended to supplement, rather than replace, the dev-error curve.