imagining syntax

← schedopt_v4

free-c1f8608e

done · peak unseen_mismatch 89.0

α schedule

α 2.978 ×66 α 2.106 ×115 α 1.495 ×119 α 1.495 ×100

direction
descending
α range
1.495–2.978
stages
4

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 1.495

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
100 54.7 ± 7.251.2 ± 4.052.5 ± 6.850.1 ± 2.9
110 56.7 ± 9.752.4 ± 5.453.6 ± 9.250.5 ± 4.1
120 59.7 ± 10.854.5 ± 7.356.1 ± 10.551.7 ± 5.3
130 64.0 ± 13.158.0 ± 10.359.8 ± 12.854.1 ± 8.2
140 68.7 ± 15.361.1 ± 11.764.2 ± 15.556.7 ± 9.8
150 74.7 ± 16.666.7 ± 13.069.9 ± 17.461.2 ± 12.0
160 80.5 ± 16.572.2 ± 14.976.0 ± 18.166.6 ± 14.3
170 84.7 ± 15.677.1 ± 13.979.8 ± 17.670.7 ± 14.1
180 87.3 ± 14.680.6 ± 13.382.7 ± 16.674.6 ± 13.5
190 89.3 ± 14.383.5 ± 12.984.9 ± 16.377.4 ± 13.4
200 91.2 ± 13.784.1 ± 14.087.8 ± 15.278.9 ± 13.7
210 93.2 ± 12.285.4 ± 13.789.9 ± 14.380.4 ± 13.9
220 93.5 ± 12.585.3 ± 15.390.2 ± 14.380.1 ± 15.0
230 93.6 ± 12.284.0 ± 15.090.4 ± 14.278.8 ± 13.9
240 94.4 ± 12.385.4 ± 15.491.4 ± 13.780.2 ± 14.4
250 94.1 ± 12.587.0 ± 15.691.9 ± 13.682.9 ± 13.8
260 94.0 ± 11.384.0 ± 17.190.7 ± 12.379.7 ± 14.6
270 94.7 ± 8.281.6 ± 17.292.0 ± 10.078.1 ± 14.2
280 95.8 ± 6.681.5 ± 16.293.0 ± 8.478.9 ± 13.4
290 96.2 ± 6.984.6 ± 17.092.8 ± 9.881.2 ± 16.4
300 94.9 ± 6.780.5 ± 11.491.4 ± 7.976.1 ± 5.5
310 97.2 ± 3.877.7 ± 18.795.4 ± 4.275.5 ± 12.3
320 97.2 ± 3.794.0 ± 4.095.3 ± 4.088.9 ± 4.1
330 98.3 ± 1.194.9 ± 1.795.5 ± 1.088.5 ± 4.8
340 99.6 ± 0.097.2 ± 0.094.8 ± 0.081.3 ± 0.0
350 98.8 ± 0.093.6 ± 0.092.4 ± 0.078.7 ± 0.0
360 99.7 ± 0.097.2 ± 0.095.1 ± 0.084.1 ± 0.0
370 99.5 ± 0.097.9 ± 0.096.3 ± 0.089.0 ± 0.0
380 99.1 ± 0.097.5 ± 0.093.1 ± 0.082.2 ± 0.0
390 99.0 ± 0.097.5 ± 0.092.6 ± 0.080.0 ± 0.0
400 99.1 ± 0.097.1 ± 0.095.1 ± 0.084.0 ± 0.0

Downloads