Generalization Gap Prediction on CIFAR-10
0.68Kendall Rank CorrelationFourier generalization measure
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Fourier generalization measure2026.06 | 0.68 | — | — | — | — | — | — | |
| PAC-Bayes flatness2026.06 | 0.613 | — | — | — | — | — | — | |
| inverse margin2026.06 | 0.473 | — | — | — | — | — | — | |
| PAC-Bayes init2026.06 | 0.429 | — | — | — | — | — | — | |
| PAC-Bayes origin2026.06 | 0.369 | — | — | — | — | — | — | |
| PAC-Bayes magnitude flatness2026.06 | 0.149 | — | — | — | — | — | — | |
| Frobenius spectral2026.06 | 0.059 | — | — | — | — | — | — | |
| # of parameters2026.06 | -0.048 | — | — | — | — | — | — | |
| PAC-Bayes magnitude init2026.06 | -0.172 | — | — | — | — | — | — | |
| Weight watcher2026.06 | -0.201 | — | — | — | — | — | — | |
| PAC-Bayes magnitude origin2026.06 | -0.236 | — | — | — | — | — | — | |
| log-prod spectral margin2026.06 | -0.256 | — | — | — | — | — | — | |
| log-sum spectral margin2026.06 | -0.256 | — | — | — | — | — | — | |
| log spectral init2026.06 | -0.266 | — | — | — | — | — | — | |
| log spectral origin2026.06 | -0.266 | — | — | — | — | — | — | |
| distance spectral init2026.06 | -0.271 | — | — | — | — | — | — | |
| log-prod spectral2026.06 | -0.315 | — | — | — | — | — | — | |
| log-sum spectral2026.06 | -0.315 | — | — | — | — | — | — | |
| path norm margin2026.06 | -0.32 | — | — | — | — | — | — | |
| path norm2026.06 | -0.369 | — | — | — | — | — | — | |
| L2-distance2026.06 | -0.414 | — | — | — | — | — | — | |
| Frobenius distance2026.06 | -0.414 | — | — | — | — | — | — | |
| L2-norm2026.06 | -0.419 | — | — | — | — | — | — | |
| parameter norm2026.06 | -0.419 | — | — | — | — | — | — | |
| log-prod Frobenius margin2026.06 | -0.429 | — | — | — | — | — | — | |
| log-sum Frobenius margin2026.06 | -0.429 | — | — | — | — | — | — | |
| log-prod Frobenius2026.06 | -0.438 | — | — | — | — | — | — | |
| log-sum Frobenius2026.06 | -0.438 | — | — | — | — | — | — | |
| 1-sharpnessTraining scenario=From scratch, Data source=Training loss landscape2023.05 | — | — | 0.75 | — | — | — | — | |
| 1-sharpnessTraining scenario=From scratch, Data source=Test loss landscape (dev data, unfair advantage)2023.05 | — | — | 0.47 | — | — | — | — | |
| adapS, path normNetwork Architecture=ResNet182026.06 | — | — | — | 0.93 | — | 9.4 | — | |
| adapS, path normNetwork Architecture=VGG132026.06 | — | — | — | 0.85 | — | 22.2 | — | |
| adapS, path normNetwork Architecture=no-BN ResNet182026.06 | — | — | — | 0.07 | — | 77.8 | — | |
| adapS, path normNetwork Architecture=no-BN VGG132026.06 | — | — | — | 0.6 | — | 63.4 | — | |
| adapS, path normNetwork Architecture=Mixed Two2026.06 | — | — | — | 0.84 | — | 19.1 | — | |
| adapS, path normNetwork Architecture=Mixed Three2026.06 | — | — | — | 0.22 | — | 59.4 | — | |
| bayesS, func normNetwork Architecture=ResNet182026.06 | — | — | — | 0.68 | — | 2.8 | — | |
| bayesS, func normNetwork Architecture=VGG132026.06 | — | — | — | 0.72 | — | 3.5 | — | |
| bayesS, func normNetwork Architecture=no-BN ResNet182026.06 | — | — | — | 0.83 | — | 1.2 | — | |
| bayesS, func normNetwork Architecture=no-BN VGG132026.06 | — | — | — | 0.7 | — | 11.4 | — | |
| bayesS, func normNetwork Architecture=Mixed Two2026.06 | — | — | — | 0.7 | — | 4 | — | |
| bayesS, func normNetwork Architecture=Mixed Three2026.06 | — | — | — | 0.74 | — | 5.1 | — | |
| Gi-scoreArchitecture=VGG, Mode=Inter-class, Layer=02021.06 | — | 3.03 | — | — | — | — | — | |
| Gi-scoreArchitecture=NiN, Mode=Inter-class, Layer=02021.06 | — | 34.34 | — | — | — | — | — | |
| HessianTraining scenario=From scratch, Data source=Training loss landscape2023.05 | — | — | 0.77 | — | — | — | — | |
| HessianTraining scenario=From scratch, Data source=Test loss landscape (dev data, unfair advantage)2023.05 | — | — | 0.58 | — | — | — | — | |
| InconsistencyTraining scenario=From scratch, Data source=Unlabeled data2023.05 | — | — | 0.17 | — | — | — | — | |
| MixupArchitecture=VGG2021.06 | — | 0.03 | — | — | — | — | — | |
| MixupArchitecture=NiN2021.06 | — | 14.18 | — | — | — | — | — | |
| Pal inter+intraArchitecture=VGG2021.06 | — | 24.84 | — | — | — | — | — | |
| Pal-scoreArchitecture=VGG, Mode=Inter-class, Layer=12021.06 | — | 7.31 | — | — | — | — | — |