Essay

Analyze the consequences of a mismatch between dev/test distribution and actual distribution

Question: Discuss why a machine learning system's dev/test set might fail to guide the development team effectively if its distribution differs from the actual distribution, and state the required action.

Sample answer: If the dev/test set distribution is not representative of the actual distribution the system needs to perform well on, the evaluation metrics will not reflect reality. This mismatch misguides the team because optimizing for the dev/test set will not improve real-world performance. When this occurs, the team must update the dev/test sets so they are more representative of the actual distribution.

Key points:

  • The actual distribution you need to do well on can be different from the dev/test sets.
  • This makes the dev/test set distribution not representative of the actual distribution.
  • The required action is to update the dev/test sets to be more representative.

Rubric: The response should explain that a non-representative dev/test set gives misleading performance signals because it differs from the actual distribution, and state that the sets must be updated to be more representative.

0

1

Updated 2026-05-27

Contributors are:

Who are from:

Tags

Machine Learning

Deep Learning

Machine Learning Strategy

Supervised Learning

Dive into Deep Learning @ D2L

Data Science

Machine Learning Yearning @ DeepLearning.AI

Related