Case Study

A team's training set accuracy is low; what should they do to diagnose the issue?

Case context: A team has trained a speech recognition system, but they notice the algorithm is performing poorly even on the training set itself, not just the dev set.

Question: Based on the source, what should this team do to understand why the algorithm is doing poorly on the training set, and what should they produce from this process?

Sample answer: The team should listen to about 100 training examples that the algorithm is performing poorly on and review them manually. From this review, they should count the errors into different categories, similar to how dev set error analysis is done, to identify the major categories of training set errors and prioritize which issues to address.

Key points:

  • Listen to ~100 poorly-performing training examples
  • Identify major categories of errors
  • Count errors in different categories
  • Process mirrors dev set error analysis

Rubric: Full credit for identifying the ~100 example review, the counting of errors by category, and the analogy to dev set error analysis.

0

1

Updated 2026-07-10

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