imagining syntax

← mirror_v1

s4_e3

done · peak unseen_mismatch 75.1

α schedule

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

direction
descending
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 = 3.0

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
50 51.7 ± 1.949.6 ± 1.049.9 ± 2.150.5 ± 1.2
100 58.6 ± 8.456.7 ± 7.254.1 ± 5.353.1 ± 4.3
150 78.4 ± 5.577.0 ± 5.264.5 ± 5.561.7 ± 5.5
200 82.4 ± 1.381.4 ± 1.765.9 ± 3.063.7 ± 4.0
250 85.8 ± 2.284.3 ± 1.667.6 ± 4.265.4 ± 2.6
300 88.2 ± 1.586.2 ± 1.970.8 ± 2.568.2 ± 3.9
350 88.4 ± 2.286.6 ± 2.170.8 ± 4.967.7 ± 5.9
400 88.1 ± 1.885.8 ± 2.469.9 ± 3.765.0 ± 5.4
450 89.4 ± 1.686.7 ± 2.171.7 ± 3.967.2 ± 4.4
500 90.8 ± 0.988.7 ± 1.572.7 ± 2.768.1 ± 4.1
550 89.9 ± 1.687.6 ± 2.573.2 ± 4.468.1 ± 4.8
600 90.9 ± 1.088.5 ± 1.674.3 ± 2.868.8 ± 4.1
650 91.9 ± 1.289.1 ± 1.876.3 ± 3.069.9 ± 5.1
700 90.8 ± 1.288.8 ± 2.475.4 ± 3.569.8 ± 6.2
750 91.7 ± 1.189.3 ± 2.176.4 ± 3.269.4 ± 5.9
800 91.6 ± 1.389.1 ± 1.875.4 ± 4.669.3 ± 7.0
850 92.5 ± 0.989.8 ± 1.877.2 ± 3.870.4 ± 5.9
900 92.7 ± 1.390.1 ± 1.976.7 ± 3.971.0 ± 6.5
950 93.0 ± 1.190.7 ± 1.679.1 ± 3.872.5 ± 5.8
1000 93.5 ± 1.490.8 ± 1.580.2 ± 3.673.4 ± 5.1
1050 93.2 ± 0.890.9 ± 1.780.2 ± 2.173.1 ± 6.6
1100 93.5 ± 0.891.2 ± 1.779.6 ± 3.373.0 ± 6.3
1150 94.7 ± 0.992.3 ± 1.382.0 ± 2.675.1 ± 5.0
1200 94.4 ± 0.891.8 ± 1.781.0 ± 3.573.0 ± 6.6

Downloads