Case Study

Diagnosing Variance in a Warehouse Package Sorter

Case context: You are evaluating a package-sorting system for a fulfillment center. Human performance on this task is nearly perfect, so the practical best error rate is about 0%. Your model achieves 2% error on the training set, 9% error on the training-dev set, and 9% error on the dev set.

Question: Based on these error metrics, what is the most likely diagnosis for the model's performance, and which comparison supports that diagnosis?

Sample answer: The most likely diagnosis is high variance. The key evidence is the jump from 2% training error to 9% training-dev error. Because the training error is already close to the ideal rate near 0%, the model does not appear to have a major bias problem. The much larger error on training-dev shows that it fits the training data better than it generalizes to new examples from the same distribution.

Key points:

  • The correct diagnosis is high variance.
  • The main evidence is the gap between 2% training error and 9% training-dev error.
  • The low training error suggests bias is not the primary issue.

Rubric: The response must explicitly name high variance and justify it by comparing the 2% training error with the 9% training-dev error.

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