imagining syntax

← refine_v1

bursty_K48

done · peak unseen_mismatch 98.5

Referenced by

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 2.1429 ×19 α 2.1429 ×19 α 2.1429 ×19 α 2.1429 ×19 α 2.1429 ×19 α 2.1429 ×19 α 2.1429 ×19 α 2.1429 ×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 1.2857 ×19 α 1.2857 ×19 α 1.2857 ×19 α 1.2857 ×19 α 1.2857 ×18 α 1.2857 ×18 α 1.2857 ×18 α 1.2857 ×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
1.0–3.0
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 51.0 ± 1.850.0 ± 1.649.8 ± 1.950.4 ± 2.0
50 50.4 ± 1.350.0 ± 1.549.9 ± 1.650.6 ± 1.8
75 52.0 ± 3.049.8 ± 1.751.6 ± 2.649.9 ± 2.1
100 55.8 ± 7.951.1 ± 3.053.8 ± 8.049.9 ± 2.3
125 66.9 ± 15.156.3 ± 12.763.8 ± 15.253.6 ± 11.3
150 81.1 ± 15.268.6 ± 12.176.8 ± 16.564.6 ± 12.4
175 93.4 ± 10.671.4 ± 15.292.5 ± 12.269.8 ± 15.3
200 99.9 ± 0.364.9 ± 21.599.9 ± 0.564.1 ± 21.2
225 100.0 ± 0.059.5 ± 29.2100.0 ± 0.059.1 ± 28.5
250 100.0 ± 0.165.8 ± 32.1100.0 ± 0.165.1 ± 32.0
275 100.0 ± 0.063.9 ± 32.9100.0 ± 0.163.1 ± 32.7
300 100.0 ± 0.066.1 ± 31.6100.0 ± 0.065.7 ± 31.6
325 99.9 ± 0.267.2 ± 36.099.8 ± 0.565.5 ± 35.4
350 99.8 ± 0.681.7 ± 21.899.5 ± 1.580.1 ± 20.8
375 99.7 ± 1.092.3 ± 16.199.2 ± 2.190.2 ± 16.4
400 99.6 ± 0.993.9 ± 13.698.6 ± 3.191.6 ± 14.6
450 99.3 ± 1.098.1 ± 2.197.5 ± 3.894.9 ± 6.7
500 100.0 ± 0.096.8 ± 12.099.8 ± 0.695.7 ± 12.5
550 100.0 ± 0.099.0 ± 3.099.9 ± 0.497.6 ± 4.2
600 100.0 ± 0.099.7 ± 0.699.9 ± 0.498.5 ± 3.3
650 99.9 ± 0.297.5 ± 6.099.0 ± 1.693.7 ± 8.3
700 99.9 ± 0.298.9 ± 1.298.9 ± 1.394.4 ± 7.0
750 99.9 ± 0.298.0 ± 2.698.8 ± 1.592.1 ± 8.7
800 100.0 ± 0.098.5 ± 3.199.8 ± 0.493.8 ± 7.8
850 100.0 ± 0.197.7 ± 2.499.4 ± 0.791.1 ± 7.9
900 100.0 ± 0.098.3 ± 2.499.7 ± 0.491.0 ± 11.3

Downloads