baseline
done · peak unseen_mismatch 90.8
α schedule
α 1.6 ×25 α 1.6 ×25 α 1.6 ×25 α 1.6 ×25 α 1.6 ×25 α 1.6 ×25 α 1.6 ×25 α 1.6 ×25 α 1.6 ×25 α 1.6 ×25 α 1.6 ×25 α 1.6 ×25 α 1.6 ×25 α 1.6 ×25 α 1.6 ×25 α 1.6 ×25
Figures
Results
accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 1.6
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 | 50.8 ± 1.7 | 49.7 ± 1.5 | 50.0 ± 1.8 | 49.8 ± 1.8 |
| 50 | 51.1 ± 1.8 | 49.9 ± 1.6 | 50.2 ± 1.8 | 50.0 ± 1.8 |
| 75 | 50.8 ± 1.6 | 49.5 ± 1.5 | 50.3 ± 1.7 | 50.0 ± 1.6 |
| 100 | 51.7 ± 2.7 | 49.5 ± 1.7 | 51.0 ± 2.4 | 49.7 ± 1.7 |
| 125 | 54.9 ± 8.9 | 49.1 ± 2.9 | 54.1 ± 9.1 | 49.2 ± 3.0 |
| 150 | 62.3 ± 16.9 | 48.8 ± 7.2 | 61.4 ± 16.9 | 48.4 ± 7.1 |
| 175 | 72.5 ± 20.2 | 51.1 ± 10.7 | 71.1 ± 20.7 | 50.1 ± 10.4 |
| 200 | 82.2 ± 20.7 | 55.3 ± 17.0 | 81.4 ± 21.2 | 54.4 ± 16.3 |
| 225 | 89.3 ± 17.9 | 59.0 ± 20.5 | 88.8 ± 18.2 | 58.0 ± 19.8 |
| 250 | 93.7 ± 14.5 | 63.5 ± 22.4 | 93.5 ± 14.7 | 62.9 ± 21.9 |
| 275 | 96.5 ± 10.9 | 69.8 ± 23.1 | 96.3 ± 11.1 | 69.0 ± 22.6 |
| 300 | 98.6 ± 6.2 | 74.7 ± 25.2 | 98.4 ± 6.5 | 73.8 ± 25.1 |
| 325 | 99.6 ± 3.0 | 79.0 ± 24.1 | 99.4 ± 3.1 | 78.1 ± 23.8 |
| 350 | 99.9 ± 0.7 | 84.7 ± 20.4 | 99.6 ± 1.2 | 83.5 ± 20.2 |
| 375 | 99.9 ± 0.2 | 88.6 ± 19.9 | 99.6 ± 1.0 | 87.5 ± 19.5 |
| 400 | 99.9 ± 0.3 | 92.2 ± 14.1 | 99.4 ± 1.3 | 90.8 ± 14.0 |