Image Classification on CIFAR-10 (test) (Accuracy vs. Initial Learning Rate Sensitivity)
86.5Test Accuracy (η_init=10^-3)Schedule-Free AdamW
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Schedule-Free AdamWBackbone=ResNet-18, Epochs=10, Baseline Type=Parameter-free / schedule-free2026.07 | 86.5 | 86.4 | — | — | — | — | |
| Vanilla AdamWBackbone=ResNet-18, Epochs=10, Base Optimizer=AdamW2026.07 | 84.3 | 80.9 | — | — | — | — | |
| Adam-InitOnly (dir.)Backbone=ResNet-18, Epochs=10, Base Optimizer=Adam, Probe=Direction-matched, Calibration=none (fixed κ = 2)2026.07 | 83.9 | 82.3 | 82.8 | 82.4 | 81.7 | 82.3 | |
| Vanilla AdamBackbone=ResNet-18, Epochs=10, Base Optimizer=Adam2026.07 | 83.4 | 81.1 | — | — | — | — | |
| Adam-InitOnlyBackbone=ResNet-18, Epochs=10, Base Optimizer=Adam2026.07 | 83.3 | 80.6 | 80.7 | 79.9 | 79.4 | 79.6 | |
| AdamW-InitOnly (dir.)Backbone=ResNet-18, Epochs=10, Base Optimizer=AdamW, Probe=Direction-matched2026.07 | 82.8 | 82.3 | 82.7 | 81.6 | 81.8 | 82.6 | |
| Adam-WatchdogBackbone=ResNet-18, Epochs=10, Base Optimizer=Adam2026.07 | 80.8 | 81.6 | 79.7 | 75.7 | 23.3 | 42.1 | |
| Adam-TrackerBackbone=ResNet-18, Epochs=10, Base Optimizer=Adam2026.07 | 80.3 | 76.6 | 75.6 | 76.5 | 76.1 | 77.2 | |
| ProdigyBackbone=ResNet-18, Epochs=10, Baseline Type=Parameter-free / schedule-free2026.07 | 10 | 52.9 | 76.8 | 79.5 | 81.3 | 82.7 |