imagining syntax

← deform_v1

end_0.7

done · peak unseen_mismatch 96.3

α schedule

α 3.0 ×19 α 3.0 ×19 α 3.0 ×19 α 3.0 ×19 α 3.0 ×19 α 3.0 ×19 α 3.0 ×19 α 3.0 ×19 α 1.0 ×19 α 1.0 ×19 α 1.0 ×19 α 1.0 ×19 α 1.0 ×19 α 1.0 ×19 α 1.0 ×19 α 1.0 ×19 α 1.85 ×19 α 1.85 ×19 α 1.85 ×19 α 1.85 ×19 α 1.85 ×19 α 1.85 ×19 α 1.85 ×19 α 1.85 ×19 α 1.0 ×19 α 1.0 ×19 α 1.0 ×19 α 1.0 ×19 α 1.0 ×19 α 1.0 ×19 α 1.0 ×19 α 1.0 ×19 α 0.7 ×19 α 0.7 ×19 α 0.7 ×19 α 0.7 ×19 α 0.7 ×18 α 0.7 ×18 α 0.7 ×18 α 0.7 ×18 α 1.0 ×18 α 1.0 ×18 α 1.0 ×18 α 1.0 ×18 α 1.0 ×18 α 1.0 ×18 α 1.0 ×18 α 1.0 ×18

direction
mixed
α range
0.7–3.0
stages
48

Figures

schedule.png
schedule.png
trajectory.png
trajectory.png

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 1.0

seen_* probed at reference α = 1.4 · unseen_* = held-out pairings, uniform (α-independent) · 1000 pairs per condition

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
25 51.4 ± 1.549.5 ± 1.649.8 ± 1.949.7 ± 2.1
50 50.7 ± 1.949.6 ± 1.450.1 ± 1.550.8 ± 1.1
75 51.9 ± 1.850.1 ± 1.950.5 ± 1.750.0 ± 1.9
100 53.8 ± 2.952.1 ± 2.951.5 ± 1.850.0 ± 1.5
125 65.8 ± 10.560.5 ± 8.159.9 ± 9.254.9 ± 5.3
150 79.3 ± 12.473.6 ± 12.872.2 ± 14.166.3 ± 12.7
175 92.1 ± 12.983.0 ± 15.889.8 ± 14.679.1 ± 16.2
200 97.0 ± 11.280.7 ± 22.797.0 ± 10.678.2 ± 21.7
225 97.3 ± 11.485.9 ± 18.997.3 ± 11.084.2 ± 18.6
250 97.4 ± 11.186.7 ± 22.297.4 ± 10.485.2 ± 22.3
275 97.5 ± 10.688.5 ± 21.497.4 ± 10.886.7 ± 21.2
300 97.6 ± 10.690.1 ± 18.297.5 ± 10.788.3 ± 18.4
325 98.4 ± 4.591.1 ± 19.396.9 ± 4.987.6 ± 18.2
350 99.3 ± 0.794.2 ± 12.497.5 ± 3.089.7 ± 12.1
375 98.9 ± 0.993.2 ± 12.496.8 ± 2.888.8 ± 12.2
400 98.5 ± 1.392.5 ± 13.094.7 ± 4.085.5 ± 12.9
450 98.3 ± 1.096.7 ± 1.393.8 ± 3.388.3 ± 4.2
500 99.9 ± 0.199.1 ± 1.499.3 ± 0.696.3 ± 2.7
550 100.0 ± 0.199.5 ± 0.699.5 ± 0.596.0 ± 2.6
600 100.0 ± 0.099.5 ± 0.799.3 ± 0.994.6 ± 6.9
650 100.0 ± 0.098.2 ± 2.799.9 ± 0.292.3 ± 9.8
700 100.0 ± 0.097.4 ± 3.6100.0 ± 0.189.8 ± 9.4
750 100.0 ± 0.093.7 ± 7.599.7 ± 0.584.1 ± 14.6
800 100.0 ± 0.097.0 ± 3.299.0 ± 1.284.8 ± 10.4
850 100.0 ± 0.097.1 ± 2.599.3 ± 0.886.5 ± 9.2
900 100.0 ± 0.095.9 ± 6.998.8 ± 1.183.7 ± 12.9

Downloads