imagining syntax

← schedopt_v4

fixed_2.6

done · peak unseen_mismatch 78.9

α schedule

α 2.6 ×400

direction
constant
α range
2.6–2.6
stages
1

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 2.6

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
100 52.8 ± 2.250.7 ± 1.750.3 ± 2.249.7 ± 1.7
110 55.0 ± 4.651.4 ± 3.351.9 ± 3.650.6 ± 2.0
120 56.3 ± 7.653.7 ± 5.853.8 ± 7.151.4 ± 4.6
130 59.7 ± 10.855.3 ± 7.556.5 ± 9.452.2 ± 5.0
140 64.3 ± 12.757.9 ± 8.060.3 ± 11.853.7 ± 5.5
150 69.5 ± 14.063.9 ± 11.364.7 ± 14.058.7 ± 9.4
160 75.3 ± 14.470.0 ± 13.769.7 ± 15.163.7 ± 12.6
170 81.1 ± 11.574.7 ± 11.474.6 ± 13.767.6 ± 12.1
180 86.0 ± 10.880.6 ± 10.678.5 ± 13.072.7 ± 11.8
190 88.6 ± 9.983.5 ± 10.281.1 ± 13.376.0 ± 11.7
200 87.9 ± 9.384.1 ± 9.479.9 ± 12.274.9 ± 10.7
210 89.7 ± 7.186.1 ± 7.880.4 ± 10.976.8 ± 10.3
220 91.1 ± 5.587.6 ± 5.881.1 ± 9.176.6 ± 7.6
230 91.9 ± 4.089.0 ± 3.381.7 ± 7.777.9 ± 6.3
240 91.9 ± 4.288.6 ± 4.680.9 ± 7.177.0 ± 5.3
250 92.9 ± 2.990.7 ± 2.781.8 ± 6.478.3 ± 5.6
260 93.2 ± 3.091.4 ± 2.582.3 ± 6.478.9 ± 5.7
270 92.7 ± 2.390.9 ± 1.979.7 ± 3.976.7 ± 3.6
280 92.7 ± 1.590.7 ± 1.380.7 ± 2.877.2 ± 3.2
290 93.6 ± 1.391.5 ± 1.281.9 ± 3.078.4 ± 3.2
300 93.8 ± 1.591.5 ± 1.382.2 ± 3.478.3 ± 2.9
310 93.4 ± 1.591.6 ± 1.681.7 ± 4.178.7 ± 3.4
320 93.5 ± 1.491.7 ± 1.381.5 ± 3.578.5 ± 2.9
330 93.9 ± 1.492.0 ± 1.681.6 ± 4.378.3 ± 3.8
340 93.6 ± 1.591.7 ± 1.680.3 ± 3.677.3 ± 3.3
350 93.8 ± 0.991.7 ± 1.879.6 ± 4.075.9 ± 4.7
360 93.7 ± 1.091.6 ± 1.480.1 ± 3.176.7 ± 3.0
370 93.8 ± 0.991.6 ± 1.380.2 ± 3.476.3 ± 2.8
380 93.6 ± 1.091.1 ± 1.379.6 ± 3.675.6 ± 3.9
390 93.8 ± 1.391.8 ± 1.779.9 ± 3.276.2 ± 3.7
400 94.0 ± 0.991.7 ± 1.478.7 ± 3.375.2 ± 3.2

Downloads