imagining syntax

← mirror_v1

s4_e5

done · peak unseen_mismatch 68.4

α schedule

α 4.0 ×50 α 4.0435 ×50 α 4.087 ×50 α 4.1304 ×50 α 4.1739 ×50 α 4.2174 ×50 α 4.2609 ×50 α 4.3043 ×50 α 4.3478 ×50 α 4.3913 ×50 α 4.4348 ×50 α 4.4783 ×50 α 4.5217 ×50 α 4.5652 ×50 α 4.6087 ×50 α 4.6522 ×50 α 4.6957 ×50 α 4.7391 ×50 α 4.7826 ×50 α 4.8261 ×50 α 4.8696 ×50 α 4.913 ×50 α 4.9565 ×50 α 5.0 ×50

direction
ascending
α range
4.0–5.0
stages
24

Figures

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

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 5.0

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 58.9 ± 8.557.1 ± 7.353.8 ± 5.452.7 ± 4.5
150 77.3 ± 4.776.7 ± 4.564.0 ± 5.661.1 ± 6.0
200 81.4 ± 1.880.5 ± 1.665.5 ± 2.963.4 ± 3.6
250 84.5 ± 1.782.7 ± 1.966.0 ± 4.564.3 ± 3.5
300 85.8 ± 1.384.6 ± 1.367.9 ± 2.465.8 ± 2.6
350 87.8 ± 2.185.9 ± 1.970.6 ± 4.367.4 ± 5.2
400 86.4 ± 2.084.1 ± 2.168.2 ± 3.964.1 ± 2.3
450 86.7 ± 1.384.7 ± 2.268.6 ± 4.764.5 ± 5.5
500 87.3 ± 1.885.2 ± 1.869.1 ± 3.164.8 ± 3.7
550 88.0 ± 1.586.0 ± 1.970.0 ± 4.067.0 ± 4.6
600 88.3 ± 2.286.5 ± 2.471.1 ± 4.767.5 ± 5.6
650 88.1 ± 1.486.0 ± 1.571.2 ± 3.766.8 ± 4.1
700 88.0 ± 1.685.8 ± 2.169.3 ± 2.765.0 ± 2.9
750 88.6 ± 1.286.4 ± 2.071.8 ± 3.868.4 ± 5.0
800 88.0 ± 1.785.6 ± 1.370.3 ± 3.567.0 ± 4.4
850 88.3 ± 1.486.8 ± 1.571.4 ± 3.467.6 ± 3.7
900 87.8 ± 1.385.6 ± 1.470.2 ± 3.666.2 ± 4.3
950 88.3 ± 1.085.9 ± 1.371.0 ± 4.167.6 ± 3.5
1000 88.4 ± 1.186.1 ± 1.370.6 ± 4.266.9 ± 3.8
1050 88.1 ± 1.585.9 ± 1.671.2 ± 3.566.7 ± 3.2
1100 89.0 ± 1.287.0 ± 1.472.9 ± 4.268.1 ± 4.3
1150 88.5 ± 1.086.5 ± 1.271.5 ± 3.167.5 ± 3.7
1200 88.4 ± 1.686.5 ± 1.270.9 ± 3.767.2 ± 2.9

Downloads