imagining syntax

← schedopt_v4

fixed_2.2

done · peak unseen_mismatch 84.6

α schedule

α 2.2 ×400

direction
constant
α range
2.2–2.2
stages
1

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 2.2

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
100 52.2 ± 2.950.3 ± 1.450.9 ± 2.649.8 ± 1.8
110 53.3 ± 4.550.7 ± 2.151.6 ± 3.850.1 ± 1.7
120 55.0 ± 9.651.4 ± 2.753.6 ± 9.350.0 ± 1.7
130 56.3 ± 11.353.5 ± 6.454.7 ± 10.451.4 ± 5.1
140 58.8 ± 12.954.1 ± 7.457.2 ± 11.251.9 ± 5.1
150 61.2 ± 13.055.7 ± 10.459.5 ± 12.352.9 ± 8.7
160 65.5 ± 15.158.9 ± 12.861.9 ± 14.756.2 ± 11.3
170 71.2 ± 16.763.8 ± 13.567.1 ± 16.659.8 ± 12.7
180 73.9 ± 17.967.6 ± 15.771.0 ± 17.564.3 ± 14.9
190 77.3 ± 18.071.2 ± 15.373.7 ± 18.666.9 ± 15.1
200 81.4 ± 17.175.6 ± 15.877.9 ± 17.471.1 ± 14.9
210 86.7 ± 14.179.8 ± 15.082.6 ± 15.574.8 ± 14.2
220 89.0 ± 13.282.9 ± 15.584.7 ± 14.377.9 ± 14.5
230 89.7 ± 14.183.7 ± 14.583.6 ± 14.877.9 ± 12.9
240 89.4 ± 13.382.6 ± 14.984.1 ± 13.275.7 ± 12.4
250 91.8 ± 11.185.2 ± 13.086.7 ± 11.878.3 ± 11.2
260 94.4 ± 6.684.5 ± 13.889.1 ± 8.377.5 ± 11.4
270 95.0 ± 5.390.1 ± 7.989.0 ± 8.481.8 ± 9.0
280 95.4 ± 3.091.1 ± 5.689.1 ± 6.982.2 ± 7.1
290 96.8 ± 2.091.5 ± 8.590.9 ± 5.482.8 ± 10.0
300 96.8 ± 1.490.2 ± 7.890.3 ± 4.781.7 ± 8.3
310 97.0 ± 1.692.3 ± 3.389.3 ± 4.982.6 ± 4.2
320 96.6 ± 1.692.8 ± 1.889.4 ± 4.382.7 ± 3.4
330 97.0 ± 1.594.1 ± 1.389.7 ± 5.584.6 ± 3.4
340 96.5 ± 1.894.3 ± 1.988.8 ± 5.584.0 ± 5.6
350 95.5 ± 0.993.1 ± 1.385.7 ± 2.280.4 ± 2.9
360 95.1 ± 0.792.6 ± 1.085.6 ± 1.980.6 ± 2.1
370 96.2 ± 0.593.2 ± 1.385.6 ± 2.880.5 ± 3.1
380 95.5 ± 0.892.8 ± 1.183.8 ± 1.777.6 ± 1.6
390 95.1 ± 1.192.7 ± 1.484.2 ± 1.878.5 ± 1.8
400 95.6 ± 1.092.9 ± 1.084.9 ± 2.278.8 ± 2.3

Downloads