imagining syntax

← slopesweep_v1

s5_e0

done · peak unseen_mismatch 93.1

Referenced by

α schedule

α 5.0 ×50 α 4.7826 ×50 α 4.5652 ×50 α 4.3478 ×50 α 4.1304 ×50 α 3.913 ×50 α 3.6957 ×50 α 3.4783 ×50 α 3.2609 ×50 α 3.0435 ×50 α 2.8261 ×50 α 2.6087 ×50 α 2.3913 ×50 α 2.1739 ×50 α 1.9565 ×50 α 1.7391 ×50 α 1.5217 ×50 α 1.3043 ×50 α 1.087 ×50 α 0.8696 ×50 α 0.6522 ×50 α 0.4348 ×50 α 0.2174 ×50 α 0.0 ×50

direction
descending
α range
0.0–5.0
stages
24

Figures

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

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 0.0

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 ± 6.957.9 ± 5.053.8 ± 3.651.3 ± 2.2
150 79.4 ± 3.477.8 ± 2.266.2 ± 5.362.6 ± 4.2
200 81.1 ± 2.580.1 ± 2.664.5 ± 4.262.9 ± 4.2
250 82.3 ± 2.681.8 ± 2.966.4 ± 4.463.8 ± 4.9
300 85.1 ± 3.384.1 ± 2.768.1 ± 6.366.0 ± 6.8
350 86.4 ± 1.585.5 ± 1.368.7 ± 3.465.9 ± 4.0
400 89.0 ± 1.488.1 ± 1.673.8 ± 3.970.6 ± 3.9
450 89.9 ± 1.488.3 ± 1.972.8 ± 4.368.7 ± 4.0
500 92.2 ± 1.989.7 ± 2.276.2 ± 5.069.8 ± 5.0
550 92.6 ± 1.390.8 ± 1.876.5 ± 3.271.3 ± 3.0
600 94.0 ± 1.691.6 ± 1.679.5 ± 4.172.6 ± 4.5
650 95.3 ± 1.393.8 ± 0.982.5 ± 4.675.7 ± 2.7
700 96.7 ± 0.994.7 ± 1.385.6 ± 3.578.4 ± 3.5
750 97.6 ± 1.096.2 ± 1.090.0 ± 3.383.3 ± 3.7
800 98.6 ± 0.696.9 ± 1.193.5 ± 3.084.5 ± 5.2
850 99.5 ± 0.298.3 ± 0.895.2 ± 2.788.1 ± 4.5
900 99.7 ± 0.298.5 ± 1.196.7 ± 2.286.8 ± 5.6
950 100.0 ± 0.199.2 ± 0.898.9 ± 1.391.7 ± 5.5
1000 100.0 ± 0.099.3 ± 1.099.8 ± 0.293.1 ± 4.4
1050 100.0 ± 0.098.7 ± 2.9100.0 ± 0.191.8 ± 10.4
1100 100.0 ± 0.090.2 ± 18.0100.0 ± 0.082.1 ± 18.7
1150 100.0 ± 0.050.4 ± 32.799.8 ± 0.345.2 ± 29.0
1200 100.0 ± 0.032.9 ± 29.0100.0 ± 0.132.1 ± 27.4

Downloads