imagining syntax

← slopesweep_v1

s5_e2

done · peak unseen_mismatch 82.5

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 51.3 ± 1.550.3 ± 1.350.3 ± 1.550.3 ± 1.8
100 61.5 ± 7.058.0 ± 5.254.0 ± 3.651.3 ± 2.2
150 79.4 ± 3.577.5 ± 2.366.2 ± 5.563.1 ± 5.2
200 80.6 ± 2.479.7 ± 2.864.1 ± 3.362.2 ± 3.5
250 81.5 ± 3.080.9 ± 2.864.8 ± 3.962.6 ± 4.3
300 84.3 ± 3.183.3 ± 3.067.9 ± 6.465.7 ± 6.5
350 84.8 ± 1.984.1 ± 1.966.4 ± 4.364.5 ± 4.0
400 87.1 ± 1.686.3 ± 2.171.1 ± 4.067.5 ± 5.7
450 87.6 ± 1.386.2 ± 2.368.9 ± 3.666.4 ± 4.7
500 88.2 ± 2.286.8 ± 2.571.1 ± 4.067.2 ± 4.7
550 88.4 ± 2.287.2 ± 2.071.4 ± 3.666.8 ± 3.5
600 89.6 ± 1.987.5 ± 2.172.8 ± 3.668.5 ± 4.1
650 89.3 ± 1.387.8 ± 1.972.8 ± 3.468.4 ± 3.6
700 91.0 ± 2.089.1 ± 1.775.6 ± 4.669.8 ± 4.0
750 91.9 ± 1.689.7 ± 1.777.3 ± 3.870.2 ± 4.6
800 92.2 ± 1.290.4 ± 1.377.7 ± 2.772.0 ± 4.2
850 92.7 ± 0.791.5 ± 1.078.2 ± 3.872.9 ± 4.6
900 93.8 ± 1.692.1 ± 1.480.6 ± 5.073.8 ± 5.9
950 95.0 ± 1.093.2 ± 1.382.2 ± 3.375.5 ± 3.6
1000 95.8 ± 0.793.7 ± 1.183.0 ± 3.475.1 ± 7.0
1050 96.3 ± 1.294.5 ± 1.384.3 ± 3.376.3 ± 5.7
1100 96.6 ± 0.694.7 ± 1.285.4 ± 3.677.3 ± 6.3
1150 97.3 ± 0.796.0 ± 1.188.2 ± 3.781.4 ± 5.3
1200 98.0 ± 0.696.7 ± 0.789.7 ± 2.882.5 ± 4.0

Downloads