p18
done · peak unseen_mismatch 70.6
α schedule
α 2.845 ×25 α 3.0295 ×25 α 3.1367 ×25 α 3.2163 ×25 α 3.2808 ×25 α 3.3357 ×25 α 3.3838 ×25 α 3.4269 ×25 α 3.466 ×25 α 3.5019 ×25 α 3.5352 ×25 α 3.5663 ×25 α 3.5955 ×25 α 3.6231 ×25 α 3.6492 ×25 α 3.6741 ×25
Figures
Results
accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 3.6741
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 | 50.9 ± 1.1 | 50.4 ± 1.6 | 50.0 ± 1.5 | 49.3 ± 1.8 |
| 50 | 51.4 ± 1.6 | 50.2 ± 1.5 | 49.7 ± 1.2 | 50.0 ± 1.8 |
| 75 | 51.9 ± 2.0 | 49.9 ± 1.8 | 50.6 ± 1.2 | 49.4 ± 1.6 |
| 100 | 54.8 ± 6.0 | 52.7 ± 2.6 | 51.9 ± 3.7 | 50.3 ± 1.6 |
| 125 | 64.1 ± 11.2 | 60.5 ± 10.2 | 57.6 ± 9.5 | 54.7 ± 7.9 |
| 150 | 74.5 ± 12.9 | 71.6 ± 12.6 | 65.1 ± 11.5 | 62.7 ± 10.4 |
| 175 | 81.6 ± 9.2 | 78.5 ± 9.0 | 69.6 ± 10.9 | 66.3 ± 9.8 |
| 200 | 86.0 ± 6.0 | 83.4 ± 5.6 | 72.3 ± 9.9 | 69.3 ± 9.0 |
| 225 | 87.2 ± 4.4 | 85.0 ± 4.4 | 72.8 ± 8.6 | 70.5 ± 8.3 |
| 250 | 87.5 ± 3.1 | 85.5 ± 2.9 | 71.9 ± 5.6 | 69.3 ± 5.7 |
| 275 | 87.9 ± 2.6 | 85.8 ± 2.2 | 71.8 ± 5.2 | 69.1 ± 5.3 |
| 300 | 88.5 ± 2.8 | 86.3 ± 2.0 | 72.2 ± 5.4 | 69.7 ± 4.8 |
| 325 | 88.6 ± 2.6 | 86.5 ± 2.2 | 71.5 ± 5.0 | 69.2 ± 5.1 |
| 350 | 89.0 ± 2.4 | 87.2 ± 2.5 | 72.9 ± 4.8 | 70.6 ± 4.6 |
| 375 | 89.1 ± 2.4 | 87.0 ± 2.1 | 72.2 ± 4.4 | 69.6 ± 4.4 |
| 400 | 89.6 ± 2.1 | 87.6 ± 1.5 | 72.8 ± 4.1 | 69.9 ± 4.4 |