Learning to Defer on Synthetic benchmark (test)
28.1Test True RiskAugmented comp-sum surrogate
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Augmented comp-sum surrogateTraining size (n)=5000, Network architecture=Two-layer MLP (32, 32) with ReLU, Surrogate type=augmented comp-sum, tau=12026.03 | 28.1 | 0.1 | 49.6 | 99.3 | |
| SeparatedTraining size (n)=5000, Network architecture=Two-layer MLP (32, 32) with ReLU, Surrogate type=exact binary logistic2026.03 | 33.4 | 5.4 | 62.5 | 58.4 | |
| L2DTraining size (n)=5000, Network architecture=Two-layer MLP (32, 32) with ReLU2026.03 | 34.2 | 6.2 | 0 | 48.8 | |
| Random route, no adviceTraining size (n)=5000, Network architecture=Two-layer MLP (32, 32) with ReLU2026.03 | 43.2 | 15.2 | 0 | 24.9 | |
| Random (j, k)Training size (n)=5000, Network architecture=Two-layer MLP (32, 32) with ReLU2026.03 | 55 | 26.9 | 50.1 | 24.9 | |
| Learned route, random adviceTraining size (n)=5000, Network architecture=Two-layer MLP (32, 32) with ReLU2026.03 | 66.1 | 38.1 | 49.6 | 24.5 |