Case context: A drone-photo app first locates a bird in an image and crops that region. A second model assigns the bird species. For one test photo, the system predicts "not snowy owl" (y=0), but the correct label is y=1. The team wants to know whether the main weakness is the bird-locating stage or the species classifier.

Question: Explain how to apply error analysis by parts to this example so the team can identify which stage caused the mistake, and say how the result should guide later work.

Sample answer: Inspect the failing example stage by stage. First, check whether the bird locator found the bird and produced a usable crop. Next, check whether the species model received that crop and still assigned the wrong label. Use those checks to decide whether the failure came from the first stage, the second stage, or both. Then repeat the same review on a larger set of mistaken cases, count which stage is most often at fault, and focus improvement effort on that stage.

Key points:

• Check the bird-locating stage on the specific failed example
• Check the species-classification stage on the specific failed example
• Assign the error to the stage or stages that actually failed
• Review many wrong predictions to see which stage causes most of them
• Direct the next round of improvement toward the most common source of error

Rubric: Full credit explains checking both stages for the single error, assigning responsibility to one or both stages, and using a wider sample of errors to decide where to concentrate future work.