imagining syntax

← schedopt_v4

fixed_2.4

done · peak unseen_mismatch 83.0

α schedule

α 2.4 ×400

direction
constant
α range
2.4–2.4
stages
1

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 2.4

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
100 52.7 ± 2.750.6 ± 1.150.6 ± 2.050.1 ± 1.9
110 54.5 ± 5.751.4 ± 2.851.7 ± 4.750.4 ± 2.3
120 55.8 ± 7.252.5 ± 4.353.6 ± 7.150.2 ± 2.9
130 57.7 ± 10.654.3 ± 7.455.1 ± 10.051.9 ± 6.0
140 62.5 ± 11.857.0 ± 9.659.5 ± 10.953.4 ± 7.7
150 67.8 ± 13.359.3 ± 10.564.8 ± 12.355.7 ± 8.5
160 74.9 ± 15.165.8 ± 12.570.7 ± 15.061.7 ± 11.0
170 79.6 ± 14.772.0 ± 13.274.8 ± 14.967.1 ± 12.2
180 81.8 ± 14.975.2 ± 12.578.1 ± 14.770.2 ± 11.0
190 87.4 ± 13.180.4 ± 12.482.4 ± 14.274.7 ± 11.8
200 90.0 ± 12.384.8 ± 12.084.4 ± 15.078.6 ± 11.8
210 91.7 ± 10.587.3 ± 10.586.2 ± 13.580.8 ± 11.5
220 91.9 ± 9.888.7 ± 9.885.6 ± 13.180.7 ± 11.0
230 92.6 ± 7.187.1 ± 10.185.4 ± 9.678.9 ± 9.6
240 92.2 ± 5.788.3 ± 7.683.8 ± 7.978.1 ± 6.6
250 93.4 ± 4.489.9 ± 6.485.1 ± 6.279.9 ± 4.4
260 94.0 ± 2.990.9 ± 3.685.4 ± 5.280.4 ± 2.4
270 94.5 ± 2.392.4 ± 1.885.2 ± 5.880.8 ± 4.3
280 95.4 ± 2.093.8 ± 2.586.1 ± 6.182.9 ± 6.4
290 95.2 ± 1.593.2 ± 1.985.0 ± 4.281.8 ± 5.0
300 95.4 ± 1.793.5 ± 2.385.7 ± 4.883.0 ± 5.3
310 94.8 ± 1.193.2 ± 1.783.9 ± 3.581.3 ± 3.1
320 94.7 ± 1.692.8 ± 1.584.5 ± 3.980.7 ± 3.9
330 94.6 ± 1.292.9 ± 1.182.4 ± 3.279.6 ± 3.9
340 94.9 ± 1.093.3 ± 1.682.6 ± 3.478.5 ± 4.4
350 94.7 ± 0.893.0 ± 1.082.3 ± 3.178.4 ± 4.2
360 95.4 ± 0.892.9 ± 1.083.4 ± 3.079.7 ± 3.9
370 95.0 ± 1.092.7 ± 1.583.3 ± 3.679.4 ± 3.9
380 94.4 ± 0.792.6 ± 1.381.5 ± 3.177.5 ± 3.0
390 94.9 ± 0.993.0 ± 1.482.0 ± 3.678.3 ± 3.7
400 95.0 ± 1.093.4 ± 1.281.2 ± 3.978.0 ± 4.8

Downloads