T200
done · peak unseen_mismatch 50.4
α schedule
α 5.0 ×5 α 5.0 ×5 α 5.0 ×5 α 5.0 ×5 α 5.0 ×5 α 5.0 ×5 α 5.0 ×5 α 5.0 ×5 α 0.7 ×4 α 0.7 ×4 α 0.7 ×4 α 0.7 ×4 α 0.7 ×4 α 0.7 ×4 α 0.7 ×4 α 0.7 ×4 α 3.1429 ×4 α 3.1429 ×4 α 3.1429 ×4 α 3.1429 ×4 α 3.1429 ×4 α 3.1429 ×4 α 3.1429 ×4 α 3.1429 ×4 α 0.7 ×4 α 0.7 ×4 α 0.7 ×4 α 0.7 ×4 α 0.7 ×4 α 0.7 ×4 α 0.7 ×4 α 0.7 ×4 α 1.2857 ×4 α 1.2857 ×4 α 1.2857 ×4 α 1.2857 ×4 α 1.2857 ×4 α 1.2857 ×4 α 1.2857 ×4 α 1.2857 ×4 α 0.7 ×4 α 0.7 ×4 α 0.7 ×4 α 0.7 ×4 α 0.7 ×4 α 0.7 ×4 α 0.7 ×4 α 0.7 ×4
Figures
Results
accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 0.7
seen_* probed at reference α = 1.4 · unseen_* = held-out pairings, uniform (α-independent) · 1000 pairs per condition
| iteration | seen_match | seen_mismatch | unseen_match | unseen_mismatch |
|---|---|---|---|---|
| 25 | 51.5 ± 1.5 | 49.9 ± 1.3 | 49.8 ± 1.9 | 49.6 ± 2.2 |
| 50 | 50.5 ± 1.5 | 49.6 ± 1.2 | 50.1 ± 1.6 | 50.4 ± 1.5 |
| 75 | 51.2 ± 1.9 | 49.9 ± 1.8 | 50.6 ± 1.9 | 49.9 ± 2.3 |
| 100 | 51.8 ± 2.1 | 50.2 ± 1.4 | 51.1 ± 2.1 | 49.1 ± 1.6 |
| 125 | 59.1 ± 10.3 | 49.3 ± 5.3 | 57.6 ± 10.5 | 48.8 ± 4.8 |
| 150 | 74.2 ± 22.0 | 48.9 ± 12.4 | 73.0 ± 22.9 | 49.7 ± 13.1 |
| 175 | 84.6 ± 18.4 | 46.5 ± 18.2 | 83.7 ± 19.6 | 46.0 ± 18.0 |
| 200 | 92.7 ± 14.7 | 41.6 ± 24.4 | 92.3 ± 14.9 | 41.7 ± 24.8 |