Case Study

Planning a prelaunch evaluation set for a pet-recognition app

Case context: You are building a mobile app that identifies birds in user photos, but the app has not yet been released. Since no real users exist yet, you do not have user-uploaded data available for creating development and test sets.

Question: Using the idea of estimating the future development/test data distribution, what should you do to build the dev and test sets, and what is one concrete way to gather data?

Sample answer: You should try to match the data you expect after launch as closely as possible. One practical step is to ask volunteers to photograph birds with their phones and share the images, so you can use those examples as a proxy for the initial dev and test sets.

Key points:

  • Dev and test sets should reflect the expected postlaunch data distribution.
  • When real user data is unavailable, proxy data must be collected.
  • Having volunteers send phone photos of birds is a suitable approximation.

Rubric: The response must state that the future user data distribution should be approximated and propose the specific action of asking volunteers to take and share mobile phone photos of birds.

0

1

Updated 2026-08-12

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