imagining syntax

← slopesweep_v1

s4_e2

done · peak unseen_mismatch 80.1

α schedule

α 4.0 ×50 α 3.913 ×50 α 3.8261 ×50 α 3.7391 ×50 α 3.6522 ×50 α 3.5652 ×50 α 3.4783 ×50 α 3.3913 ×50 α 3.3043 ×50 α 3.2174 ×50 α 3.1304 ×50 α 3.0435 ×50 α 2.9565 ×50 α 2.8696 ×50 α 2.7826 ×50 α 2.6957 ×50 α 2.6087 ×50 α 2.5217 ×50 α 2.4348 ×50 α 2.3478 ×50 α 2.2609 ×50 α 2.1739 ×50 α 2.087 ×50 α 2.0 ×50

direction
descending
α range
2.0–4.0
stages
24

Figures

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

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 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 51.4 ± 1.250.2 ± 1.349.9 ± 1.650.4 ± 1.8
100 55.9 ± 4.653.8 ± 4.252.2 ± 2.750.5 ± 2.1
150 81.7 ± 5.578.0 ± 6.370.5 ± 8.565.3 ± 7.3
200 84.6 ± 3.683.3 ± 3.569.9 ± 7.167.0 ± 6.0
250 85.1 ± 1.984.7 ± 2.170.0 ± 4.966.8 ± 4.9
300 87.3 ± 1.787.2 ± 1.972.3 ± 4.469.5 ± 5.4
350 89.0 ± 1.587.8 ± 1.473.2 ± 4.269.9 ± 4.5
400 90.3 ± 1.989.0 ± 2.475.9 ± 6.172.1 ± 7.4
450 90.4 ± 1.889.0 ± 2.074.9 ± 4.670.3 ± 6.0
500 91.6 ± 2.089.4 ± 2.476.3 ± 4.671.1 ± 6.2
550 91.5 ± 1.489.6 ± 1.476.3 ± 4.470.3 ± 4.0
600 92.8 ± 1.690.8 ± 1.779.0 ± 4.672.2 ± 4.9
650 93.1 ± 1.890.9 ± 1.878.6 ± 5.270.7 ± 5.6
700 93.3 ± 1.491.4 ± 1.281.5 ± 3.573.6 ± 3.9
750 94.1 ± 1.291.4 ± 1.081.8 ± 3.873.7 ± 4.5
800 94.5 ± 1.192.0 ± 1.382.1 ± 5.074.2 ± 4.9
850 95.0 ± 0.993.0 ± 1.383.6 ± 3.474.1 ± 5.0
900 95.7 ± 1.193.3 ± 1.285.7 ± 3.676.2 ± 5.1
950 96.5 ± 1.093.6 ± 1.587.0 ± 3.876.5 ± 5.6
1000 96.7 ± 1.094.4 ± 0.887.8 ± 1.978.3 ± 4.8
1050 97.3 ± 0.794.5 ± 1.387.2 ± 3.076.7 ± 6.0
1100 97.6 ± 0.695.5 ± 0.988.9 ± 3.179.6 ± 4.7
1150 97.8 ± 0.995.2 ± 1.089.9 ± 2.478.9 ± 5.9
1200 98.3 ± 0.696.2 ± 1.291.6 ± 2.680.1 ± 6.7

Downloads