Required Quality of the Verification Target
Question: Answer in one to three sentences: What must the reference output y* satisfy, and what is not necessary?
Sample answer: y* should be a result that scores better than the model's current output under the chosen evaluation rule. It does not need to be the single best possible answer, only an answer that is better for comparison.
Key points:
- y* must beat the present model output.
- y* does not have to be globally optimal.
Rubric: The response must include both the comparison to the current output and the fact that global optimality is unnecessary.
0
1
Tags
Machine Learning
Deep Learning
Supervised Learning
Dive into Deep Learning @ D2L
Data Science
Machine Learning Strategy
Machine Learning Yearning @ DeepLearning.AI
Related
Which property should a human reference trajectory have in optimization verification?
A verification comparison output must be provably optimal.
The candidate output must be _____ when compared with the current model's output.
Match each optimization verification term with its meaning.
Order the steps for using a human demonstration that is better than the current policy output.
Explain why a better reference is enough for this comparison test.
Can a better human route serve as the reference in Optimization Verification?
Required Quality of the Verification Target
If a human-produced solution is strong but not perfect, what can it still be used for?
A stronger-than-baseline result can still help decide what to improve.