imagining syntax

← schedopt_v4

fixed_0.8

done · peak unseen_mismatch 50.3

α schedule

α 0.8 ×400

direction
constant
α range
0.8–0.8
stages
1

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 0.8

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
100 50.6 ± 2.149.9 ± 1.150.0 ± 1.950.3 ± 1.9
110 51.3 ± 1.749.5 ± 1.350.5 ± 1.649.6 ± 1.9
120 50.8 ± 1.649.6 ± 1.751.0 ± 2.049.3 ± 1.4
130 51.3 ± 1.949.5 ± 1.951.5 ± 2.049.5 ± 1.6
140 51.1 ± 3.449.4 ± 2.351.5 ± 2.749.3 ± 1.9
150 51.5 ± 5.148.8 ± 2.752.3 ± 4.348.6 ± 2.8
160 53.1 ± 7.348.3 ± 3.252.8 ± 7.848.4 ± 2.8
170 54.7 ± 10.948.3 ± 4.054.3 ± 11.648.0 ± 3.7
180 56.1 ± 12.946.3 ± 9.455.7 ± 12.946.0 ± 9.3
190 58.4 ± 14.244.5 ± 12.058.0 ± 14.344.5 ± 11.9
200 59.9 ± 15.644.9 ± 11.160.2 ± 15.744.5 ± 11.2
210 63.1 ± 17.546.5 ± 7.662.7 ± 17.846.4 ± 6.8
220 68.6 ± 19.444.4 ± 12.367.3 ± 19.844.2 ± 11.9
230 73.9 ± 21.346.4 ± 13.273.5 ± 21.546.4 ± 13.0
240 78.1 ± 21.140.2 ± 16.078.0 ± 21.340.9 ± 16.3
250 82.8 ± 19.938.5 ± 12.382.3 ± 20.537.7 ± 12.2
260 86.7 ± 19.031.4 ± 17.186.5 ± 18.831.8 ± 16.2
270 90.2 ± 18.029.2 ± 18.790.1 ± 18.129.2 ± 18.9
280 91.3 ± 17.227.2 ± 17.491.4 ± 17.126.9 ± 17.5
290 92.6 ± 15.728.6 ± 19.193.1 ± 14.827.8 ± 18.7
300 94.9 ± 13.124.7 ± 22.195.0 ± 12.924.0 ± 21.8
310 96.8 ± 10.924.2 ± 19.896.8 ± 11.024.2 ± 20.2
320 97.6 ± 9.528.5 ± 23.697.4 ± 10.028.3 ± 23.2
330 98.0 ± 8.622.7 ± 20.598.0 ± 8.522.5 ± 20.3
340 98.7 ± 5.521.4 ± 19.798.8 ± 5.220.1 ± 19.1
350 99.5 ± 2.215.7 ± 18.499.5 ± 2.315.0 ± 17.7
360 100.0 ± 0.018.2 ± 17.9100.0 ± 0.017.5 ± 17.5
370 100.0 ± 0.019.8 ± 20.1100.0 ± 0.019.3 ± 20.2
380 100.0 ± 0.019.8 ± 20.3100.0 ± 0.019.7 ± 20.5
390 100.0 ± 0.120.0 ± 21.099.9 ± 0.219.7 ± 20.9
400 100.0 ± 0.017.4 ± 23.5100.0 ± 0.017.2 ± 23.5

Downloads