p21
done · peak unseen_mismatch 90.7
α schedule
α 1.5908 ×25 α 1.6881 ×25 α 1.7669 ×25 α 1.8371 ×25 α 1.9017 ×25 α 1.9623 ×25 α 2.0199 ×25 α 2.075 ×25 α 2.128 ×25 α 2.1794 ×25 α 2.2291 ×25 α 2.2776 ×25 α 2.3249 ×25 α 2.3711 ×25 α 2.4164 ×25 α 2.4608 ×25
Figures
Results
accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 2.4608
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 | 50.5 ± 1.5 | 50.0 ± 1.6 | 49.8 ± 1.7 | 49.7 ± 2.0 |
| 50 | 50.9 ± 1.5 | 49.9 ± 1.5 | 50.1 ± 1.7 | 49.7 ± 1.5 |
| 75 | 50.7 ± 1.5 | 49.6 ± 1.8 | 50.4 ± 1.3 | 49.7 ± 1.8 |
| 100 | 50.5 ± 2.5 | 50.1 ± 1.7 | 50.1 ± 1.4 | 50.0 ± 1.3 |
| 125 | 52.3 ± 2.8 | 49.6 ± 2.0 | 51.2 ± 3.2 | 49.3 ± 2.1 |
| 150 | 56.8 ± 11.9 | 51.0 ± 4.8 | 55.4 ± 11.4 | 49.2 ± 4.1 |
| 175 | 67.9 ± 17.2 | 56.2 ± 9.0 | 66.4 ± 16.7 | 54.3 ± 7.7 |
| 200 | 77.4 ± 20.8 | 61.2 ± 16.6 | 76.1 ± 20.8 | 60.0 ± 16.0 |
| 225 | 83.3 ± 21.3 | 72.6 ± 20.4 | 82.2 ± 21.4 | 71.2 ± 19.4 |
| 250 | 88.1 ± 17.9 | 80.7 ± 18.4 | 86.3 ± 18.6 | 78.9 ± 17.9 |
| 275 | 93.2 ± 12.4 | 88.7 ± 13.3 | 91.4 ± 13.3 | 85.8 ± 14.1 |
| 300 | 95.5 ± 10.8 | 93.0 ± 10.3 | 93.7 ± 10.9 | 89.8 ± 10.7 |
| 325 | 96.6 ± 8.1 | 93.8 ± 10.1 | 94.0 ± 8.9 | 90.7 ± 9.9 |
| 350 | 97.1 ± 5.0 | 93.5 ± 9.7 | 93.7 ± 5.9 | 89.2 ± 9.5 |
| 375 | 97.6 ± 1.7 | 95.5 ± 2.9 | 93.8 ± 4.2 | 90.3 ± 5.4 |
| 400 | 97.3 ± 1.8 | 95.9 ± 2.3 | 92.8 ± 4.9 | 90.1 ± 6.2 |