imagining syntax

← schedopt_v4

fixed_2.7

done · peak unseen_mismatch 78.4

α schedule

α 2.7 ×400

direction
constant
α range
2.7–2.7
stages
1

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 2.7

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
100 53.6 ± 3.851.3 ± 1.750.8 ± 3.250.2 ± 2.1
110 55.9 ± 6.652.2 ± 3.652.7 ± 5.450.5 ± 2.7
120 58.7 ± 9.855.0 ± 6.655.4 ± 8.951.7 ± 4.9
130 61.6 ± 11.557.8 ± 8.257.8 ± 11.253.0 ± 6.1
140 66.4 ± 12.960.8 ± 8.861.4 ± 13.055.6 ± 7.0
150 72.3 ± 13.266.5 ± 11.566.5 ± 13.659.9 ± 9.8
160 77.2 ± 13.771.1 ± 13.770.8 ± 14.564.6 ± 12.2
170 81.8 ± 12.276.9 ± 11.674.5 ± 14.469.1 ± 11.7
180 83.3 ± 11.279.2 ± 10.575.3 ± 12.869.7 ± 10.5
190 87.6 ± 8.582.7 ± 9.178.7 ± 11.573.0 ± 9.6
200 89.2 ± 7.385.5 ± 7.880.6 ± 10.476.0 ± 9.1
210 89.8 ± 6.286.4 ± 6.580.2 ± 8.875.6 ± 7.7
220 90.3 ± 4.987.2 ± 5.679.9 ± 8.175.5 ± 6.4
230 91.3 ± 3.988.8 ± 4.180.3 ± 7.276.4 ± 5.9
240 91.5 ± 3.289.6 ± 3.379.9 ± 6.677.0 ± 5.5
250 92.5 ± 2.790.6 ± 2.481.7 ± 6.377.8 ± 5.2
260 92.8 ± 2.890.7 ± 2.881.7 ± 6.378.4 ± 5.3
270 92.2 ± 2.790.7 ± 2.880.5 ± 5.976.9 ± 5.5
280 92.6 ± 1.790.6 ± 1.780.3 ± 3.976.0 ± 3.2
290 92.9 ± 1.590.7 ± 1.680.3 ± 3.776.7 ± 3.4
300 93.2 ± 1.691.2 ± 1.581.1 ± 3.476.9 ± 3.4
310 93.1 ± 1.491.2 ± 1.780.5 ± 4.276.9 ± 3.3
320 93.1 ± 1.490.9 ± 1.680.6 ± 3.576.1 ± 3.9
330 93.6 ± 1.491.4 ± 1.780.3 ± 4.576.2 ± 3.9
340 93.5 ± 1.391.5 ± 1.580.4 ± 3.975.9 ± 3.6
350 93.1 ± 1.491.1 ± 1.578.5 ± 4.275.0 ± 4.0
360 93.6 ± 1.491.8 ± 1.180.4 ± 3.176.3 ± 3.3
370 93.4 ± 1.291.3 ± 1.379.4 ± 3.975.1 ± 4.1
380 93.2 ± 1.391.1 ± 1.578.8 ± 3.674.3 ± 4.7
390 93.8 ± 1.191.7 ± 1.379.7 ± 4.275.0 ± 4.5
400 93.6 ± 1.391.3 ± 1.678.0 ± 5.274.1 ± 4.0

Downloads