imagining syntax

← schedopt_v5_confirm

flat_1_8

done · peak unseen_mismatch 80.4

Referenced by

α schedule

α 1.8 ×400

direction
constant
α range
1.8–1.8
stages
1

Figures

schedule.png
schedule.png

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 1.8

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
100 51.6 ± 3.449.8 ± 1.650.7 ± 3.249.7 ± 1.7
110 52.5 ± 4.950.0 ± 1.851.6 ± 4.849.7 ± 1.9
120 54.0 ± 7.049.8 ± 2.553.1 ± 6.949.5 ± 2.4
130 57.1 ± 10.250.3 ± 3.955.4 ± 10.149.6 ± 3.5
140 59.8 ± 13.150.5 ± 5.458.3 ± 13.349.8 ± 4.8
150 63.8 ± 16.052.2 ± 7.062.2 ± 16.151.1 ± 6.5
160 68.1 ± 18.454.0 ± 9.466.4 ± 18.752.7 ± 8.9
170 73.1 ± 19.957.3 ± 12.571.4 ± 20.555.8 ± 11.8
180 76.8 ± 20.058.9 ± 15.275.3 ± 20.957.4 ± 14.5
190 81.3 ± 19.560.8 ± 16.879.8 ± 20.259.1 ± 15.8
200 84.6 ± 18.962.7 ± 16.683.4 ± 19.661.1 ± 15.9
210 87.4 ± 18.063.9 ± 18.586.2 ± 18.962.4 ± 17.9
220 89.1 ± 16.863.6 ± 18.288.1 ± 17.562.0 ± 17.4
230 91.5 ± 15.468.5 ± 19.490.6 ± 16.266.6 ± 18.6
240 92.5 ± 14.569.5 ± 20.291.4 ± 15.467.6 ± 19.4
250 94.1 ± 12.370.5 ± 19.793.1 ± 13.368.3 ± 19.0
260 95.3 ± 11.072.6 ± 20.794.3 ± 12.770.9 ± 20.3
270 96.4 ± 8.572.9 ± 18.095.5 ± 9.870.8 ± 17.1
280 97.4 ± 7.178.0 ± 17.896.6 ± 8.475.8 ± 17.2
290 97.8 ± 6.374.9 ± 16.296.9 ± 8.672.8 ± 15.2
300 98.5 ± 5.075.7 ± 17.497.8 ± 6.274.1 ± 16.3
310 99.1 ± 2.978.6 ± 18.698.4 ± 5.476.3 ± 18.0
320 99.2 ± 2.676.2 ± 19.498.6 ± 4.474.4 ± 18.8
330 99.3 ± 1.671.3 ± 16.898.4 ± 4.069.1 ± 15.0
340 99.4 ± 1.469.4 ± 19.698.6 ± 3.167.6 ± 17.4
350 99.4 ± 1.368.9 ± 18.198.8 ± 2.767.0 ± 16.3
360 99.5 ± 1.174.5 ± 20.698.3 ± 3.372.7 ± 19.1
370 99.2 ± 1.569.6 ± 24.997.9 ± 3.466.8 ± 21.9
380 99.3 ± 1.571.8 ± 18.897.8 ± 4.369.1 ± 14.4
390 99.3 ± 1.479.3 ± 18.597.7 ± 4.776.7 ± 14.8
400 99.2 ± 1.482.9 ± 16.797.2 ± 5.080.4 ± 13.9

Downloads