Essay

How to isolate the source of an error in a three-stage workflow

Question: In a pipeline with three stages, A, B, and C, describe a practical method for identifying which stage caused a mistake seen on the dev set.

Sample answer: A common way to localize the problem is to test the later stages with an idealized input. First, replace A’s output with the correct output that A should have produced, then run B and C. If the final result becomes correct, A is the likely source of the error. If the result is still wrong, keep A fixed and instead replace B’s output with its correct version, then run C. If that makes the overall output correct, the error came from B. If the output remains incorrect even after giving B a correct input, then the remaining stage, C, is responsible.

Key points:

  • Start by substituting the correct output for stage A and evaluate B and C.
  • If the pipeline is fixed, attribute the mistake to A.
  • If not, substitute the correct output for stage B and evaluate C.
  • If that fixes the result, attribute the mistake to B; otherwise, blame C.

Rubric: The response should explain a step-by-step attribution procedure that begins with A, then moves to B, and assigns the error to C if the problem persists after giving A and B correct inputs.

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