imagining syntax

← schedopt_v4

free-d7d5317e

done · peak unseen_mismatch 89.7

α schedule

α 2.904 ×83 α 1.832 ×14 α 3.12 ×20 α 1.497 ×49 α 2.879 ×66 α 1.625 ×43 α 2.942 ×25 α 2.942 ×100

direction
mixed
α range
1.497–3.12
stages
8

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 2.942

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
100 55.0 ± 5.751.5 ± 3.052.5 ± 5.449.8 ± 2.4
110 58.6 ± 9.953.6 ± 6.155.2 ± 9.051.3 ± 4.2
120 62.1 ± 11.756.2 ± 7.458.3 ± 12.052.8 ± 5.9
130 66.0 ± 14.058.3 ± 9.261.2 ± 14.354.7 ± 8.2
140 71.3 ± 15.760.9 ± 11.966.8 ± 16.656.6 ± 10.9
150 77.1 ± 16.465.5 ± 14.373.1 ± 17.861.5 ± 13.7
160 82.8 ± 16.370.4 ± 14.679.5 ± 17.466.1 ± 13.9
170 88.6 ± 14.376.2 ± 15.586.0 ± 15.872.6 ± 15.2
180 91.3 ± 11.779.8 ± 13.788.7 ± 12.975.3 ± 12.9
190 93.6 ± 7.384.4 ± 11.590.3 ± 9.479.5 ± 11.3
200 94.6 ± 5.887.8 ± 8.691.1 ± 8.282.3 ± 8.7
210 95.2 ± 3.790.0 ± 5.690.5 ± 6.783.1 ± 6.3
220 95.0 ± 3.290.8 ± 4.689.6 ± 6.783.4 ± 6.0
230 94.8 ± 2.890.4 ± 7.088.6 ± 6.382.2 ± 7.5
240 96.1 ± 2.292.1 ± 6.491.0 ± 4.684.7 ± 6.3
250 97.8 ± 1.295.0 ± 3.594.2 ± 3.189.3 ± 3.8
260 97.9 ± 1.594.1 ± 7.393.7 ± 4.487.5 ± 7.3
270 98.4 ± 0.995.9 ± 3.494.5 ± 3.089.7 ± 4.2
280 97.7 ± 1.394.6 ± 2.492.5 ± 3.985.5 ± 4.7
290 94.7 ± 2.292.1 ± 2.585.9 ± 6.380.4 ± 7.2
300 93.7 ± 0.790.9 ± 1.385.3 ± 3.478.9 ± 3.5
310 93.5 ± 0.990.8 ± 1.585.0 ± 1.479.1 ± 3.0
320 93.1 ± 0.990.2 ± 1.482.0 ± 3.076.2 ± 3.4
330 92.7 ± 1.390.4 ± 0.981.6 ± 2.876.0 ± 2.4
340 93.2 ± 0.790.6 ± 1.081.6 ± 2.476.0 ± 4.1
350 93.2 ± 1.090.6 ± 1.182.2 ± 2.476.1 ± 2.2
360 92.8 ± 0.990.4 ± 1.182.3 ± 2.876.3 ± 3.2
370 92.6 ± 1.089.9 ± 1.181.3 ± 1.676.4 ± 2.7
380 93.5 ± 0.790.6 ± 1.681.5 ± 1.075.7 ± 2.7
390 93.2 ± 1.290.8 ± 1.379.9 ± 2.774.0 ± 2.2
400 92.5 ± 1.290.5 ± 1.180.0 ± 2.774.5 ± 2.7

Downloads