imagining syntax

← schedopt_v4

fixed_2.5

done · peak unseen_mismatch 80.7

α schedule

α 2.5 ×400

direction
constant
α range
2.5–2.5
stages
1

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 2.5

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 ± 2.850.8 ± 1.350.4 ± 2.450.2 ± 2.0
110 54.3 ± 3.751.1 ± 2.551.8 ± 2.950.3 ± 1.7
120 55.6 ± 5.352.4 ± 3.553.0 ± 4.150.1 ± 2.2
130 58.1 ± 9.354.2 ± 4.955.1 ± 9.151.3 ± 2.5
140 62.9 ± 12.056.4 ± 7.059.1 ± 11.753.1 ± 4.8
150 66.7 ± 14.259.8 ± 10.763.0 ± 13.555.9 ± 8.4
160 73.4 ± 15.665.8 ± 12.868.6 ± 15.360.6 ± 10.8
170 77.0 ± 15.570.4 ± 13.271.6 ± 16.265.0 ± 11.6
180 81.6 ± 14.876.2 ± 13.676.1 ± 15.669.4 ± 13.4
190 84.6 ± 12.277.5 ± 12.578.1 ± 13.970.3 ± 11.1
200 87.9 ± 10.982.7 ± 10.880.7 ± 13.575.5 ± 10.7
210 90.9 ± 9.186.5 ± 9.883.2 ± 13.578.4 ± 12.0
220 91.0 ± 9.487.8 ± 9.883.0 ± 12.579.1 ± 11.1
230 91.7 ± 6.688.2 ± 7.481.8 ± 9.678.2 ± 8.9
240 92.0 ± 5.289.7 ± 5.682.2 ± 8.178.6 ± 7.4
250 93.2 ± 4.790.4 ± 5.983.0 ± 6.779.1 ± 5.9
260 93.7 ± 3.291.3 ± 3.783.6 ± 5.180.1 ± 4.2
270 93.7 ± 2.291.7 ± 2.582.8 ± 4.879.2 ± 3.7
280 94.0 ± 2.191.9 ± 2.383.3 ± 4.579.9 ± 5.0
290 94.0 ± 1.891.7 ± 2.282.4 ± 3.879.3 ± 3.4
300 94.2 ± 1.391.4 ± 1.882.7 ± 4.579.3 ± 4.0
310 93.9 ± 1.391.6 ± 1.782.3 ± 3.379.2 ± 2.9
320 94.3 ± 1.292.2 ± 1.684.0 ± 2.980.7 ± 2.6
330 94.4 ± 1.292.6 ± 1.183.1 ± 3.680.0 ± 3.1
340 94.0 ± 1.392.0 ± 1.982.1 ± 3.878.6 ± 4.5
350 94.4 ± 1.092.1 ± 1.982.4 ± 4.178.5 ± 5.1
360 93.8 ± 1.091.8 ± 1.781.7 ± 2.677.8 ± 3.6
370 94.1 ± 1.292.2 ± 1.782.2 ± 2.978.3 ± 3.6
380 94.2 ± 1.592.0 ± 1.782.0 ± 4.277.8 ± 3.4
390 94.4 ± 0.992.4 ± 1.481.7 ± 4.377.4 ± 3.5
400 93.9 ± 1.091.8 ± 1.579.5 ± 4.276.7 ± 3.6

Downloads