Case Study

Choosing an Error Check Before Expanding the Dataset

Case context: A company is developing a model that sorts drone photos into crop-health categories. One manager wants the team to spend several weeks gathering and labeling thousands of additional images right away. Another engineer argues that the team should first inspect the current model’s bias and variance.

Question: Using the idea that error comes mainly from two sources, explain why the engineer’s suggestion should come before a large data-collection effort.

Sample answer: The engineer’s suggestion is the better starting point because bias and variance account for the main kinds of machine learning error. Measuring them tells the team whether the current model is failing because it is too simple or because it changes too much across different samples. That diagnosis matters because extra training examples help only in some situations, so checking these error patterns first avoids spending effort on a step that may not address the real problem.

Key points:

  • State that bias and variance are the two main sources of error.
  • Explain that examining them helps determine whether more training data is likely to help.
  • Conclude that this check prevents an unproductive labeling effort.

Rubric: Student must explain that bias and variance are the two main error sources, and that examining them is needed to decide whether collecting more data is a worthwhile use of time.

0

1

Updated 2026-08-12

Contributors are:

Who are from:

Tags

D2L

Dive into Deep Learning @ D2L

Machine Learning

Deep Learning

Supervised Learning

Data Science

Machine Learning Strategy

Machine Learning Yearning @ DeepLearning.AI

Related