imagining syntax

← schedopt_v4

fixed_2.3

done · peak unseen_mismatch 84.3

α schedule

α 2.3 ×400

direction
constant
α range
2.3–2.3
stages
1

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 2.3

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
100 53.4 ± 4.750.6 ± 1.551.3 ± 4.449.9 ± 1.8
110 54.7 ± 7.050.9 ± 1.952.4 ± 5.850.1 ± 1.7
120 56.1 ± 10.252.9 ± 5.554.2 ± 10.350.9 ± 4.6
130 57.3 ± 11.654.4 ± 8.155.7 ± 10.652.4 ± 7.6
140 61.9 ± 13.656.4 ± 9.659.4 ± 12.354.1 ± 8.5
150 65.8 ± 15.259.3 ± 13.364.0 ± 14.156.3 ± 12.2
160 70.5 ± 14.862.1 ± 11.566.5 ± 14.458.5 ± 10.2
170 76.2 ± 16.169.8 ± 13.972.0 ± 16.165.0 ± 12.8
180 80.7 ± 16.373.3 ± 13.776.9 ± 16.467.7 ± 12.7
190 84.5 ± 16.277.8 ± 14.280.2 ± 17.272.8 ± 13.8
200 86.6 ± 14.279.3 ± 12.482.1 ± 16.074.4 ± 12.3
210 90.2 ± 12.484.2 ± 12.785.6 ± 15.578.8 ± 13.4
220 90.1 ± 12.685.2 ± 12.284.2 ± 15.179.2 ± 12.6
230 90.0 ± 12.585.7 ± 12.384.1 ± 14.479.0 ± 11.7
240 91.5 ± 11.588.3 ± 12.485.2 ± 12.781.5 ± 11.3
250 90.7 ± 12.588.8 ± 12.383.9 ± 11.380.2 ± 10.9
260 91.8 ± 9.489.9 ± 10.084.6 ± 9.580.9 ± 10.1
270 91.9 ± 6.790.2 ± 7.083.4 ± 5.880.4 ± 5.9
280 93.0 ± 5.490.9 ± 6.285.2 ± 3.781.6 ± 4.9
290 94.5 ± 2.692.7 ± 4.186.8 ± 2.683.0 ± 2.7
300 95.1 ± 2.993.2 ± 4.386.9 ± 3.084.3 ± 4.1
310 96.1 ± 1.494.0 ± 1.786.8 ± 3.184.0 ± 3.0
320 95.7 ± 1.593.3 ± 2.287.2 ± 3.083.8 ± 3.4
330 96.2 ± 0.994.3 ± 0.986.8 ± 4.383.7 ± 4.0
340 96.4 ± 1.494.1 ± 1.286.7 ± 5.283.6 ± 5.5
350 96.4 ± 0.994.8 ± 1.587.6 ± 2.883.6 ± 3.4
360 95.1 ± 0.893.8 ± 1.584.8 ± 1.682.0 ± 2.1
370 95.3 ± 1.094.4 ± 1.485.1 ± 2.181.8 ± 3.6
380 95.1 ± 0.994.2 ± 1.483.9 ± 2.981.5 ± 3.6
390 95.2 ± 1.294.1 ± 1.984.8 ± 3.281.3 ± 4.3
400 95.0 ± 0.893.8 ± 1.083.3 ± 3.180.0 ± 3.9

Downloads