Case Study

Diagnosing a cat recognizer with high training and dev error rates.

Case context: You are building a cat recognizer with a target error rate of 5%. Currently, the model achieves a 15% error rate on the training set and a 16% error rate on the dev set. Your colleague proposes collecting 10,000 more training images to resolve this issue.

Question: Based on the error rates and target, diagnose the primary problem, evaluate the proposed solution, and state what the first step should be to improve performance.

Sample answer: The model is suffering from avoidable bias because the training error of 15% is much higher than the target error of 5%. Collecting 10,000 more training images will not help because adding training data resolves variance (since training and dev error are already close at 15% and 16%, variance is low) but has no significant effect on bias. The first step should be to focus on improving the algorithm's performance on the current training set, rather than gathering more data.

Key points:

  • Identify avoidable bias as the primary issue due to the 10% gap between training error (15%) and target error (5%).
  • Explain that collecting more training data is an ineffective solution because it addresses variance, not bias.
  • Recommend focusing first on improving the model's performance on the training set.

Rubric: Diagnose avoidable bias based on the 15% training error vs 5% target error. Evaluate the proposal to collect more data as ineffective because it addresses variance rather than bias. Recommend focusing on improving training-set performance first.

0

1

Updated 2026-05-26

Contributors are:

Who are from:

Tags

Machine Learning

Deep Learning

Supervised Learning

Dive into Deep Learning @ D2L

Data Science

Machine Learning Strategy

Related