Essay

Explain what poor final performance means when the tuning set and final set do or do not match.

Question: A spam filter scores well on the tuning set but poorly on the final evaluation set. Explain how you would interpret this result and what you would do differently in two cases: (1) the tuning and final sets were drawn from the same population, and (2) they came from different populations.

Sample answer: When the tuning and final sets come from the same population, the diagnosis is straightforward: the model has likely adapted too closely to the tuning set. In that case, adding more tuning examples is a natural remedy because it reduces the chance that quirks of the tuning set are driving model choices.

When the tuning and final sets come from different populations, the situation is less clear. The gap could reflect adaptation to the tuning set, a genuinely tougher final set, or simply the best performance the current approach can realistically achieve. Because several explanations fit the evidence, the next step is not obvious from these scores alone.

Key points:

  • If the tuning and final sets come from the same population, strong tuning performance with weak final performance points to overfitting on the tuning set.
  • In that same-population case, a sensible fix is to collect additional tuning examples.
  • If the two sets come from different populations, the diagnosis is not definitive.
  • Under different populations, possible explanations include tuning-set overfitting, a more difficult final set, or the model already being near its practical limit.

Rubric: To receive full credit, the answer must identify that: 1. When the two sets come from the same population, the issue is overfitting to the tuning set and the remedy is to get more tuning data. 2. When the two sets come from different populations, the diagnosis is uncertain. 3. The possible explanations in the different-population case include overfitting to the tuning set, the final set being harder, or the method performing about as well as it can.

0

1

Updated 2026-08-12

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

Related