imagining syntax

← mirror_v1

s3_e4

done · peak unseen_mismatch 72.1

α schedule

α 3.0 ×50 α 3.0435 ×50 α 3.087 ×50 α 3.1304 ×50 α 3.1739 ×50 α 3.2174 ×50 α 3.2609 ×50 α 3.3043 ×50 α 3.3478 ×50 α 3.3913 ×50 α 3.4348 ×50 α 3.4783 ×50 α 3.5217 ×50 α 3.5652 ×50 α 3.6087 ×50 α 3.6522 ×50 α 3.6957 ×50 α 3.7391 ×50 α 3.7826 ×50 α 3.8261 ×50 α 3.8696 ×50 α 3.913 ×50 α 3.9565 ×50 α 4.0 ×50

direction
ascending
α range
3.0–4.0
stages
24

Figures

schedule.png
schedule.png
trajectory.png
trajectory.png

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 4.0

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
50 51.3 ± 2.249.6 ± 0.850.0 ± 1.750.3 ± 1.6
100 56.7 ± 6.553.9 ± 4.754.0 ± 5.850.8 ± 2.7
150 76.8 ± 13.374.0 ± 12.268.7 ± 13.264.2 ± 11.2
200 88.2 ± 3.586.7 ± 3.275.9 ± 6.172.1 ± 6.5
250 89.0 ± 2.187.6 ± 1.774.5 ± 3.971.5 ± 3.4
300 90.6 ± 1.489.0 ± 0.975.2 ± 3.571.8 ± 2.6
350 91.1 ± 1.989.5 ± 2.175.6 ± 4.371.7 ± 4.1
400 90.3 ± 2.488.2 ± 1.673.5 ± 5.569.0 ± 4.1
450 91.4 ± 1.389.3 ± 1.475.6 ± 5.070.8 ± 5.4
500 91.5 ± 0.689.5 ± 1.274.5 ± 2.769.8 ± 3.5
550 92.0 ± 1.189.9 ± 0.976.3 ± 3.071.9 ± 3.3
600 91.4 ± 1.189.3 ± 1.274.8 ± 3.670.2 ± 4.5
650 92.4 ± 1.289.8 ± 2.176.1 ± 5.070.4 ± 5.4
700 91.7 ± 0.889.7 ± 1.775.2 ± 3.070.8 ± 4.7
750 92.3 ± 1.489.9 ± 1.576.1 ± 4.371.2 ± 6.0
800 91.4 ± 1.589.9 ± 1.674.6 ± 5.870.3 ± 6.3
850 91.4 ± 0.989.8 ± 1.576.3 ± 3.971.5 ± 4.2
900 91.4 ± 1.689.7 ± 1.375.5 ± 4.471.2 ± 5.2
950 91.7 ± 1.289.5 ± 1.176.0 ± 3.870.5 ± 4.1
1000 91.5 ± 1.189.6 ± 1.275.7 ± 3.970.2 ± 4.9
1050 91.5 ± 1.289.2 ± 1.175.3 ± 4.269.5 ± 3.6
1100 91.2 ± 1.289.0 ± 1.575.8 ± 4.869.9 ± 5.5
1150 91.8 ± 1.590.0 ± 1.077.4 ± 4.671.5 ± 5.4
1200 91.8 ± 1.389.9 ± 0.974.6 ± 5.070.5 ± 5.4

Downloads