Case Study

Smoothing a Noisy Dev Set Learning Curve at Small Training Sizes

Case context: You are training a model and plotting its learning curves. At small training sizes (e.g., 10 examples), the plotted training and dev error values fluctuate wildly, making it impossible to determine if the model is experiencing high bias or high variance. You want to apply the random subset averaging technique using your original set of 100 training examples.

Question: Based on the provided case context, how should you construct the training subsets, and how will you obtain the final values to plot on the learning curve for the small training size of 10 examples?

Sample answer: To smooth the noise at the small training size of 10 examples, I should select 3 to 10 different training subsets of 10 examples each from the original 100 examples by sampling with replacement. I will then train a different model on each of these subsets and compute the training error and dev set error for each model. Finally, I will compute the average training error and average dev set error across these models and plot these average values on the learning curve.

Key points:

  • Construct 3-10 training subsets of 10 examples by sampling with replacement from the original 100 examples.
  • Train a different model on each of the constructed subsets.
  • Compute individual training and dev set errors for each of the trained models.
  • Average the training and dev errors across all models and plot the averages.

Rubric: The response must outline: 1) selecting 3-10 subsets of 10 examples using sampling with replacement from the 100 original examples, 2) training a different model on each subset, 3) calculating training and dev set error for each model, and 4) averaging and plotting the resulting training and dev errors.

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