Case Study

Design a dev-set error learning curve experiment for an algorithm with 1,000 available training examples.

Case context: An engineer has accumulated a dataset of 1,000 labeled training examples. To diagnose whether the model would benefit from more data, the engineer decides to build a learning curve that plots dev-set error against training set size.

Question: Based on the provided concepts for constructing learning curves, describe how the engineer should set up the training runs and outline what the resulting plot should display.

Sample answer: The engineer should select several training-set sizes spanning up to the total of 1,000 examples, such as 100, 200, 300, and so on. They must train separate copies of the algorithm on each of these subset sizes. After training, the dev-set error is evaluated for each copy. The engineer then plots these dev-set error values on the vertical axis (y-axis) against the corresponding training-set sizes on the horizontal axis (x-axis) to create the learning curve.

Key points:

  • Train separate copies of the algorithm on training subsets of varying sizes (e.g., 100, 200, 300, ..., 1,000).
  • Evaluate the dev-set error for each of the trained algorithm copies.
  • Plot the resulting dev-set error values against their respective training-set sizes.

Rubric: The response must detail: 1) selecting incremental subset sizes up to the maximum available (such as 100, 200, 300 up to 1,000), 2) training separate copies of the algorithm on these subsets, and 3) evaluating and plotting dev-set error versus the training-set sizes.

0

1

Updated 2026-05-27

Contributors are:

Who are from:

Tags

Machine Learning

Deep Learning

Supervised Learning

Dive into Deep Learning @ D2L

Data Science

Machine Learning Strategy

Machine Learning Yearning @ DeepLearning.AI

Related