Essay

How should you partition dev set errors using the Optimization Verification test?

Question: Describe the process of applying the Optimization Verification test across all errors in a dev set. Detail how each error is classified based on comparing the scores of the correct output SS^* and the algorithm's output SoutS_{out}.

Sample answer: To apply the test in practice, you examine every error in your dev set. For each error, you compare ScoreA(S)Score_A(S^*) with ScoreA(Sout)Score_A(S_{out}). If ScoreA(S)>ScoreA(Sout)Score_A(S^*) > Score_A(S_{out}) holds, the error is classified as an optimization-algorithm error. If this inequality does not hold (i.e., ScoreA(S)ScoreA(Sout)Score_A(S^*) \le Score_A(S_{out})), the error is classified as a scoring-function error.

Key points:

  • Examine all errors in the dev set rather than focusing on a single example.
  • For each error, evaluate the inequality condition ScoreA(S)>ScoreA(Sout)Score_A(S^*) > Score_A(S_{out}).
  • Mark errors where the inequality holds as optimization-algorithm errors.
  • Count errors where the inequality does not hold as scoring-function errors.

Rubric: The answer should describe examining all dev set errors, comparing ScoreA(S)Score_A(S^*) and ScoreA(Sout)Score_A(S_{out}), and mapping the results to optimization-algorithm errors (when ScoreA(S)>ScoreA(Sout)Score_A(S^*) > Score_A(S_{out})) and scoring-function errors (when ScoreA(S)ScoreA(Sout)Score_A(S^*) \le Score_A(S_{out})).

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