Conformal Prediction Performance Metrics on CIFAR10
96.93Top-1 AccuracyLabel Distance
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Label DistanceConformal domain=Probability space, Distance metric=Cosine distance, Model=ResNet-502026.05 | 96.93 | 99.87 | 1.144 | |
| Margin DistanceConformal domain=Logit space, Distance metric=Cosine distance, Model=ResNet-502026.05 | 96.93 | 99.87 | 1.108 | |
| Mean DistanceConformal domain=Logit space, Distance metric=Euclidean distance, Model=ResNet-502026.05 | 96.93 | 99.87 | 1.082 | |
| APS (U=0.001)Conformal domain=Probability space, Distance metric=N/A, Model=ResNet-502026.05 | 96.93 | 99.87 | 1.193 | |
| RAPS (λ = 0.2, k = 1, U=0.001)Conformal domain=Probability space, Distance metric=N/A, Model=ResNet-502026.05 | 96.93 | 99.87 | 1.35 | |
| SAPS (λ = 0.2, U=0.001)Conformal domain=Probability space, Distance metric=N/A, Model=ResNet-502026.05 | 96.93 | 99.87 | 1.263 | |
| Gradient (lr=0.1, 100 steps)Conformal domain=Feature space, Distance metric=Euclidean distance, Model=ResNet-502026.05 | 96.93 | 99.87 | 1.238 | |
| Fast GradientConformal domain=Feature space, Distance metric=Euclidean distance, Model=ResNet-502026.05 | 96.93 | 99.87 | 1.442 |