s2_e2
done · peak unseen_mismatch 94.8
α schedule
α 2.0 ×50 α 2.0 ×50 α 2.0 ×50 α 2.0 ×50 α 2.0 ×50 α 2.0 ×50 α 2.0 ×50 α 2.0 ×50 α 2.0 ×50 α 2.0 ×50 α 2.0 ×50 α 2.0 ×50 α 2.0 ×50 α 2.0 ×50 α 2.0 ×50 α 2.0 ×50 α 2.0 ×50 α 2.0 ×50 α 2.0 ×50 α 2.0 ×50 α 2.0 ×50 α 2.0 ×50 α 2.0 ×50 α 2.0 ×50
Figures
Results
accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 2.0
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 |
|---|---|---|---|---|
| 50 | 50.5 ± 1.9 | 49.9 ± 0.5 | 50.1 ± 1.9 | 50.1 ± 1.5 |
| 100 | 51.4 ± 2.2 | 50.6 ± 1.9 | 50.9 ± 1.8 | 48.9 ± 1.1 |
| 150 | 63.6 ± 17.2 | 58.1 ± 11.5 | 62.7 ± 16.4 | 56.0 ± 9.3 |
| 200 | 83.3 ± 19.9 | 77.6 ± 18.4 | 81.7 ± 20.7 | 74.7 ± 17.8 |
| 250 | 96.9 ± 7.3 | 93.9 ± 8.5 | 95.8 ± 8.4 | 91.1 ± 9.9 |
| 300 | 99.3 ± 0.8 | 98.0 ± 1.6 | 97.3 ± 2.8 | 94.8 ± 3.4 |
| 350 | 98.5 ± 1.2 | 96.9 ± 1.4 | 95.1 ± 4.2 | 91.5 ± 4.9 |
| 400 | 97.7 ± 1.1 | 95.9 ± 1.3 | 92.3 ± 4.4 | 88.1 ± 4.7 |
| 450 | 97.5 ± 0.7 | 95.7 ± 1.4 | 91.7 ± 2.9 | 87.5 ± 4.5 |
| 500 | 96.8 ± 0.5 | 95.5 ± 1.1 | 89.2 ± 1.8 | 85.3 ± 1.6 |
| 550 | 97.2 ± 0.9 | 95.8 ± 1.4 | 90.3 ± 2.9 | 86.2 ± 3.3 |
| 600 | 97.0 ± 1.1 | 95.3 ± 1.7 | 88.7 ± 3.4 | 84.1 ± 3.7 |
| 650 | 97.2 ± 0.8 | 95.4 ± 1.3 | 89.1 ± 2.8 | 83.5 ± 4.5 |
| 700 | 97.2 ± 1.1 | 95.6 ± 1.6 | 88.1 ± 3.8 | 84.0 ± 4.4 |
| 750 | 97.6 ± 0.9 | 95.8 ± 1.2 | 90.1 ± 3.3 | 84.8 ± 4.3 |
| 800 | 97.5 ± 0.8 | 95.7 ± 1.1 | 89.5 ± 3.2 | 84.6 ± 4.3 |
| 850 | 97.4 ± 1.0 | 96.0 ± 1.3 | 89.6 ± 5.0 | 84.9 ± 5.4 |
| 900 | 97.9 ± 1.1 | 96.7 ± 1.0 | 89.8 ± 3.5 | 85.4 ± 5.1 |
| 950 | 98.1 ± 0.6 | 96.2 ± 0.7 | 89.4 ± 3.6 | 85.2 ± 3.3 |
| 1000 | 97.9 ± 0.8 | 96.6 ± 0.9 | 89.7 ± 3.0 | 85.1 ± 3.0 |
| 1050 | 98.3 ± 0.4 | 96.8 ± 1.4 | 91.2 ± 2.8 | 86.2 ± 3.6 |
| 1100 | 98.3 ± 0.6 | 96.7 ± 1.0 | 90.9 ± 2.7 | 85.3 ± 3.5 |
| 1150 | 98.6 ± 0.6 | 97.2 ± 0.9 | 91.6 ± 2.0 | 86.9 ± 4.0 |
| 1200 | 98.2 ± 0.5 | 96.8 ± 0.7 | 90.9 ± 3.8 | 84.3 ± 4.7 |