Case Study

Deciding whether to build a manual review set for satellite burst detection.

Case context: A machine learning team is building a detector for rare burst noise in satellite telemetry. The bursts last only a few milliseconds and are so domain-specific that field technicians cannot reliably tell whether a flagged segment is truly a burst or a normal fluctuation. The team is considering whether to spend time creating an Eyeball dev set of flagged examples for human inspection.

Question: Given that humans cannot reliably judge these examples, what should the team do about the Eyeball dev set, and why?

Sample answer: They should skip the Eyeball dev set. If people cannot confidently tell whether a case is correct, then manual inspection will not help them understand why the model is wrong, so the set would not be a useful debugging tool.

Key points:

  • Do not create the Eyeball dev set.
  • Human reviewers cannot perform the task reliably.
  • Without human judgment, error analysis will not explain model mistakes well.

Rubric: The answer should recommend skipping the Eyeball dev set and explain that manual review is not useful when humans cannot reliably do the underlying task or diagnose the model's errors.

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