Case Study

Choose evaluation strategy for a crop disease app.

Case context: A startup is building a mobile system that identifies leaf diseases from photos taken by farmers. The team has a large archive of high-quality greenhouse images and a smaller set of field photos captured in varied lighting, backgrounds, and camera conditions. An engineer suggests spending significant time on cross-domain transfer methods so the model can be trained on greenhouse images and then adapted to field photos.

Question: For the startup's immediate product goals, should the team prioritize that transfer-focused approach? Explain your choice.

Sample answer: No. For near-term product work, the team should make the evaluation data match the real target use case: field photos from farmers. Transfer methods can be useful in some situations, but the first priority is to build dev and test sets from the target population so progress is measured on the data that matters most. That usually helps the team move faster and make clearer decisions about model improvements.

Key points:

  • The immediate goal is building a useful product, not exploring a broad research direction.
  • The dev and test sets should reflect the field-photo distribution.
  • Cross-domain transfer is not the main first step for this situation.
  • Matching evaluation data to the target setting improves team efficiency and decision-making.

Rubric: The response should recommend against prioritizing the transfer-focused approach for the immediate goal and instead advise using dev and test sets drawn from the target field-photo distribution to improve team efficiency.

0

1

Updated 2026-08-12

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