imagining syntax

← schedopt_v4

fixed_2.8

done · peak unseen_mismatch 76.6

α schedule

α 2.8 ×400

direction
constant
α range
2.8–2.8
stages
1

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 2.8

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
100 54.5 ± 3.851.4 ± 2.451.2 ± 3.449.7 ± 1.8
110 56.0 ± 5.752.2 ± 3.252.3 ± 4.950.4 ± 2.1
120 58.6 ± 9.455.1 ± 6.055.1 ± 8.251.5 ± 4.0
130 62.1 ± 11.558.0 ± 8.058.5 ± 10.753.8 ± 5.7
140 67.1 ± 13.062.4 ± 9.262.8 ± 12.857.2 ± 7.0
150 73.3 ± 13.968.3 ± 10.967.1 ± 14.061.4 ± 9.3
160 78.2 ± 14.174.1 ± 12.670.9 ± 14.365.8 ± 11.6
170 81.5 ± 12.877.1 ± 11.873.3 ± 14.068.4 ± 11.3
180 84.5 ± 11.080.8 ± 9.876.2 ± 13.270.9 ± 11.2
190 87.3 ± 8.283.5 ± 7.677.8 ± 12.072.5 ± 9.6
200 88.6 ± 7.486.0 ± 7.878.6 ± 10.674.8 ± 10.0
210 89.9 ± 4.587.0 ± 4.978.8 ± 8.074.9 ± 6.9
220 90.2 ± 3.887.5 ± 4.278.6 ± 7.075.1 ± 5.3
230 91.1 ± 3.288.6 ± 3.079.5 ± 6.776.1 ± 5.4
240 90.8 ± 2.989.2 ± 2.678.0 ± 6.675.3 ± 5.5
250 91.7 ± 2.489.7 ± 2.278.8 ± 6.175.9 ± 4.7
260 91.7 ± 2.490.0 ± 2.079.1 ± 4.876.0 ± 3.9
270 92.0 ± 2.490.1 ± 2.179.1 ± 4.876.3 ± 3.1
280 92.1 ± 1.990.4 ± 1.979.4 ± 4.376.6 ± 3.5
290 92.4 ± 1.690.6 ± 1.579.3 ± 3.876.2 ± 3.2
300 92.8 ± 1.590.7 ± 1.580.2 ± 3.376.5 ± 2.7
310 92.1 ± 1.390.3 ± 1.578.9 ± 3.275.5 ± 2.4
320 92.1 ± 1.490.4 ± 1.878.7 ± 3.874.9 ± 3.6
330 92.5 ± 1.690.6 ± 1.877.9 ± 3.974.5 ± 3.4
340 92.7 ± 1.790.9 ± 1.378.3 ± 4.674.6 ± 3.6
350 92.6 ± 1.390.7 ± 1.578.4 ± 4.273.9 ± 4.3
360 92.8 ± 1.790.8 ± 1.978.7 ± 4.174.6 ± 4.9
370 92.9 ± 1.491.0 ± 1.878.4 ± 3.674.3 ± 4.1
380 92.7 ± 1.390.3 ± 1.777.6 ± 4.173.6 ± 4.4
390 93.3 ± 1.290.9 ± 1.578.7 ± 4.374.2 ± 3.7
400 93.1 ± 1.690.7 ± 1.577.4 ± 4.673.0 ± 3.5

Downloads