imagining syntax

← schedopt_v5_confirm

free-d56d6dda

done · peak unseen_mismatch 98.0

α schedule

α 4.4 ×10 α 4.9 ×10 α 4.1 ×10 α 5.0 ×10 α 4.5 ×20 α 4.8 ×10 α 4.2 ×10 α 4.7 ×10 α 4.5 ×10 α 5.0 ×10 α 4.9 ×20 α 4.4 ×10 α 4.5 ×10 α 4.3 ×10 α 4.4 ×10 α 1.9 ×10 α 0.2 ×20 α 0.0 ×10 α 0.2 ×10 α 2.7 ×10 α 4.3 ×20 α 4.2 ×10 α 1.6 ×10 α 0.5 ×10 α 0.1 ×10 α 0.7 ×110

direction
mixed
α range
0.0–5.0
stages
26

Figures

schedule.png
schedule.png

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 0.7

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
100 60.1 ± 7.456.6 ± 6.453.7 ± 5.451.5 ± 3.6
110 65.6 ± 8.861.7 ± 7.756.5 ± 7.253.6 ± 5.0
120 70.5 ± 8.966.9 ± 7.759.5 ± 8.056.1 ± 5.4
130 74.6 ± 8.471.5 ± 7.361.7 ± 8.758.5 ± 6.4
140 77.5 ± 6.774.6 ± 6.063.2 ± 8.160.2 ± 6.5
150 79.1 ± 5.476.6 ± 4.963.9 ± 7.361.1 ± 6.2
160 80.2 ± 4.677.8 ± 4.264.4 ± 7.061.5 ± 5.9
170 80.8 ± 3.878.6 ± 3.564.4 ± 6.861.8 ± 5.6
180 83.8 ± 3.981.6 ± 3.667.8 ± 7.364.9 ± 6.4
190 88.9 ± 5.586.5 ± 5.578.1 ± 9.974.5 ± 9.5
200 95.0 ± 5.592.7 ± 6.490.8 ± 8.887.2 ± 9.4
210 98.2 ± 3.094.3 ± 8.296.9 ± 4.592.0 ± 9.2
220 99.3 ± 1.684.5 ± 18.498.9 ± 2.282.5 ± 18.4
230 99.9 ± 0.287.0 ± 11.399.6 ± 0.484.0 ± 12.2
240 99.4 ± 0.795.9 ± 2.998.7 ± 1.792.1 ± 4.8
250 95.2 ± 2.192.4 ± 1.888.9 ± 4.783.9 ± 3.6
260 91.8 ± 1.889.5 ± 1.182.6 ± 4.178.0 ± 2.7
270 96.7 ± 1.294.4 ± 2.091.7 ± 1.787.1 ± 2.8
280 99.9 ± 0.298.9 ± 1.299.3 ± 0.998.0 ± 2.4

Downloads