Is it possible to increase the precision to 100%?
Lele-Xie opened this issue · comments
Hi~ Recently, I have read some paper of yours (OpenMix, FMFP) about MisD or OOD. I am curious about whether it is possible to increase the precision to 100% by sacrificing some recall rate after using some related confidence techniques.
Thanks for your interest in our papers. For a specific dataset/model, you can plot the Accuracy-Rejection curves (fig.7 in OpenMix paper) or the Risk-Coverage curves, and find the rejection rate where the corresponding accuracy is 100% (or the risk is 0). To achieve exact 100% accuracy, the rejection rate might be high, and this depends on the dataset and model.
Hello~ Our recent work SURE:SUrvey REcipes for building reliable and robust deep networks(CVPR2024) also focus on Failure Prediction(MisD) task.