Case context: You are developing a machine learning model, and its current dev-set error is unacceptably high. You want to determine if investing months into gathering more training data will actually help reduce the error.

Question: How should you use a learning curve in this situation to inform your decision?

Sample answer: You should plot a learning curve by calculating and graphing the dev-set error for various subsets of your current training data. By observing the trend of the dev-set error as the number of training examples increases, you can extrapolate whether the curve is still trending downwards (meaning more data will likely help) or if it has flattened out (meaning more data will not significantly reduce the error).

Key points:

• Plot dev-set error against training set size
• Observe the trend as training examples increase
• Extrapolate to make a decision about gathering data

Rubric: The answer should identify plotting the dev-set error against the number of training examples to observe the trend and evaluate the benefit of more data.