Essay

How Dev-Set Error Rate Affects Eyeball Dev Set Size

Question: In error analysis, explain how a classifier's error rate affects the size of the Eyeball dev set. Include a numerical example showing how to estimate the needed number of examples to produce enough mistakes for manual review.

Sample answer: The relationship is inverse: as the classifier gets better and its error rate drops, you need a larger Eyeball dev set to collect enough wrong predictions for useful inspection. A common target is on the order of 100 errors so that patterns become visible. For example, if a classifier has a 4% error rate and you want about 100 mistakes, then the dev set should contain about 2,500 examples because 0.04 × 2,500 = 100.

Key points:

  • Error rate and Eyeball dev set size move in opposite directions.
  • The purpose is to gather enough misclassified examples for analysis, often around 100.
  • The example calculation should show that 4% of 2,500 equals 100.

Rubric: The answer must clearly state the inverse relationship, explain why the goal is to collect roughly 100 errors for inspection, and correctly work through the 4% example calculation.

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