Question: Explain how repeated random subsampling can make learning curves easier to interpret when the training set is small. Describe how the subsamples are chosen, how many models are trained, which errors are measured, and how the final plot is produced.

Sample answer: When only a small amount of training data is available, a single learning curve can bounce around a lot because one unlucky sample may not represent the full data distribution. A practical way to smooth this out is to draw several random training subsets of the same size, often around 5 to 8 subsets, from the full training pool. Each subset is selected independently, using sampling with replacement if the goal is to mimic repeated draws from the population.

A separate model is trained on each subset. For every trained model, two quantities are measured: the error on its own training subset and the error on the development set. After collecting those values across all runs, the practitioner averages the training errors and averages the development errors. The two averaged series are then plotted against training-set size to reveal a cleaner learning trend.

Key points:

• Draw several random subsets of the same small size from the available training data.
• Use sampling with replacement when repeated draws are needed.
• Train one model per subset.
• Measure training error and development-set error for each model.
• Average the errors across runs before plotting the learning curves.

Rubric: A strong response explains how the subsets are formed, notes that multiple separate models are trained, identifies both training and development errors as the measured quantities, and states that averaging those errors produces a smoother curve.