imagining syntax

← mirror_v1

s5_e2

done · peak unseen_mismatch 83.2

α schedule

α 5.0 ×50 α 4.8696 ×50 α 4.7391 ×50 α 4.6087 ×50 α 4.4783 ×50 α 4.3478 ×50 α 4.2174 ×50 α 4.087 ×50 α 3.9565 ×50 α 3.8261 ×50 α 3.6957 ×50 α 3.5652 ×50 α 3.4348 ×50 α 3.3043 ×50 α 3.1739 ×50 α 3.0435 ×50 α 2.913 ×50 α 2.7826 ×50 α 2.6522 ×50 α 2.5217 ×50 α 2.3913 ×50 α 2.2609 ×50 α 2.1304 ×50 α 2.0 ×50

direction
descending
α range
2.0–5.0
stages
24

Figures

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

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 2.0

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
50 51.8 ± 2.149.9 ± 1.250.2 ± 1.850.0 ± 1.4
100 60.2 ± 7.458.5 ± 7.154.0 ± 4.952.6 ± 4.5
150 75.7 ± 5.674.1 ± 5.861.5 ± 4.458.9 ± 4.4
200 80.2 ± 2.178.6 ± 1.463.0 ± 3.161.3 ± 3.0
250 82.4 ± 2.381.4 ± 2.164.6 ± 4.162.6 ± 2.9
300 84.3 ± 2.183.0 ± 2.066.7 ± 4.564.5 ± 3.4
350 87.1 ± 2.285.4 ± 2.569.2 ± 3.166.0 ± 4.0
400 86.5 ± 2.284.5 ± 2.468.6 ± 4.664.0 ± 5.1
450 87.0 ± 1.685.0 ± 2.468.8 ± 5.463.2 ± 4.6
500 88.6 ± 1.186.1 ± 2.270.1 ± 5.165.1 ± 5.3
550 89.2 ± 1.087.2 ± 2.171.9 ± 5.467.4 ± 6.2
600 89.9 ± 1.487.6 ± 2.172.3 ± 3.567.8 ± 4.7
650 91.2 ± 1.789.0 ± 2.474.5 ± 6.069.1 ± 6.8
700 91.0 ± 1.288.6 ± 2.274.0 ± 3.968.7 ± 5.3
750 92.3 ± 1.190.2 ± 2.278.3 ± 3.971.3 ± 6.5
800 92.7 ± 1.390.4 ± 2.277.9 ± 3.871.1 ± 5.3
850 93.5 ± 0.991.1 ± 2.179.8 ± 4.972.4 ± 6.0
900 94.3 ± 1.192.3 ± 1.681.3 ± 3.275.1 ± 4.9
950 95.0 ± 1.192.8 ± 1.782.5 ± 3.975.1 ± 6.3
1000 95.5 ± 0.593.2 ± 1.483.0 ± 2.674.4 ± 5.0
1050 95.9 ± 0.994.1 ± 1.884.5 ± 3.677.5 ± 4.9
1100 96.6 ± 0.394.7 ± 1.586.1 ± 3.779.0 ± 4.8
1150 97.9 ± 0.596.3 ± 1.189.6 ± 1.982.7 ± 4.4
1200 98.4 ± 0.596.6 ± 0.991.0 ± 3.883.2 ± 6.7

Downloads