Case Study

Diagnosing Variance in a Cat-Detection Algorithm

Case context: You are evaluating a cat-detection algorithm. You know that human-level performance is nearly perfect, setting the optimal error rate at about 0%. After running your model, you observe a 1% error on the training set, a 5% error on the training-dev set, and a 5% error on the dev set.

Question: Based on the provided error metrics, what is the primary diagnosis for the algorithm's performance, and which specific error comparison justifies this conclusion?

Sample answer: The primary diagnosis is that the algorithm has high variance. This is justified by comparing the 1% training error to the 5% training-dev error. Since the 1% training error is already very close to the 0% optimal error, bias is not the main issue. The 4% increase in error when moving to the training-dev set shows the model struggles to generalize to new data from the same distribution.

Key points:

  • Primary diagnosis is high variance.
  • Justification relies on the gap between the 1% training error and 5% training-dev error.
  • Notes that the training error (1%) is close to the optimal error (~0%), ruling out high bias.

Rubric: The answer must explicitly state 'high variance' as the diagnosis and justify it by highlighting the gap between the 1% training error and 5% training-dev error.

0

1

Updated 2026-06-13

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