imagining syntax

← mirror_v1

s0_e4

done · peak unseen_mismatch 89.9

α schedule

α 0.0 ×50 α 0.1739 ×50 α 0.3478 ×50 α 0.5217 ×50 α 0.6957 ×50 α 0.8696 ×50 α 1.0435 ×50 α 1.2174 ×50 α 1.3913 ×50 α 1.5652 ×50 α 1.7391 ×50 α 1.913 ×50 α 2.087 ×50 α 2.2609 ×50 α 2.4348 ×50 α 2.6087 ×50 α 2.7826 ×50 α 2.9565 ×50 α 3.1304 ×50 α 3.3043 ×50 α 3.4783 ×50 α 3.6522 ×50 α 3.8261 ×50 α 4.0 ×50

direction
ascending
α range
0.0–4.0
stages
24

Figures

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

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 4.0

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
50 50.4 ± 1.649.9 ± 1.149.8 ± 1.949.6 ± 1.9
100 50.0 ± 1.449.9 ± 1.649.5 ± 1.049.5 ± 1.2
150 50.4 ± 2.349.3 ± 2.750.9 ± 3.249.0 ± 1.9
200 60.6 ± 16.745.5 ± 6.959.3 ± 17.045.4 ± 7.0
250 69.2 ± 23.241.7 ± 15.268.8 ± 23.241.3 ± 15.8
300 76.3 ± 24.038.5 ± 17.676.3 ± 23.840.1 ± 18.3
350 81.4 ± 23.233.6 ± 18.281.0 ± 23.533.3 ± 17.3
400 86.9 ± 20.134.7 ± 18.386.2 ± 21.235.9 ± 18.9
450 99.3 ± 2.126.4 ± 22.899.3 ± 2.228.2 ± 23.8
500 100.0 ± 0.043.0 ± 28.9100.0 ± 0.043.2 ± 28.9
550 100.0 ± 0.052.8 ± 30.1100.0 ± 0.053.9 ± 30.1
600 100.0 ± 0.063.1 ± 33.4100.0 ± 0.063.1 ± 32.4
650 100.0 ± 0.071.7 ± 32.7100.0 ± 0.171.0 ± 31.3
700 99.7 ± 0.888.2 ± 22.098.8 ± 2.886.5 ± 21.0
750 99.6 ± 0.791.5 ± 16.198.5 ± 2.389.4 ± 16.2
800 98.8 ± 1.193.2 ± 12.196.6 ± 3.389.9 ± 11.3
850 97.3 ± 1.692.3 ± 12.093.9 ± 4.487.8 ± 11.1
900 96.6 ± 1.395.3 ± 1.491.5 ± 4.389.0 ± 3.6
950 95.0 ± 1.693.9 ± 1.587.4 ± 5.984.7 ± 5.1
1000 93.4 ± 1.292.0 ± 1.583.6 ± 4.180.7 ± 4.4
1050 92.4 ± 1.591.2 ± 1.580.9 ± 4.677.9 ± 3.8
1100 91.3 ± 1.190.0 ± 1.378.2 ± 4.275.2 ± 2.9
1150 89.8 ± 1.189.1 ± 1.074.8 ± 4.672.2 ± 3.9
1200 89.2 ± 1.188.0 ± 1.273.5 ± 4.170.9 ± 3.7

Downloads