imagining syntax

← curriculum_v1

cds

done · peak unseen_mismatch 81.0

Referenced by

α schedule

α 1.46 ×150 α 1.4 ×150 α 1.44 ×150 α 1.38 ×150 α 1.37 ×150 α 1.28 ×150 α 1.23 ×150 α 1.25 ×150

direction
mixed
α range
1.23–1.46
stages
8

Figures

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

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 1.25

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
150 56.4 ± 14.648.1 ± 4.357.5 ± 14.448.4 ± 4.2
300 96.7 ± 10.030.7 ± 21.096.3 ± 11.230.3 ± 20.6
450 100.0 ± 0.034.7 ± 30.4100.0 ± 0.034.8 ± 30.2
600 100.0 ± 0.045.9 ± 30.0100.0 ± 0.046.9 ± 29.9
750 100.0 ± 0.065.9 ± 34.599.9 ± 0.464.8 ± 33.5
900 100.0 ± 0.076.1 ± 30.899.9 ± 0.274.3 ± 29.6
1050 100.0 ± 0.181.0 ± 30.399.3 ± 1.478.1 ± 28.7
1200 99.9 ± 0.185.4 ± 26.299.1 ± 1.781.0 ± 25.6

Downloads