imagining syntax

← curriculum_v1

noisy_14

done · peak unseen_mismatch 88.9

Referenced by

α schedule

α 1.4 ×150 α 1.4 ×150 α 1.4 ×150 α 1.4 ×150 α 1.4 ×150 α 1.4 ×150 α 1.4 ×150 α 1.4 ×150

direction
constant
α range
1.4–1.4
stages
8

Figures

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

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 1.4

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
150 52.5 ± 2.248.9 ± 2.250.5 ± 2.049.9 ± 1.4
300 99.7 ± 0.830.7 ± 16.399.6 ± 0.930.5 ± 16.0
450 100.0 ± 0.038.9 ± 20.5100.0 ± 0.038.6 ± 20.2
600 100.0 ± 0.066.9 ± 25.4100.0 ± 0.066.4 ± 24.7
750 99.9 ± 0.171.1 ± 26.199.5 ± 1.370.5 ± 25.8
900 100.0 ± 0.174.9 ± 20.799.4 ± 1.774.0 ± 20.2
1050 99.8 ± 0.393.5 ± 7.998.3 ± 2.388.9 ± 8.8
1200 99.5 ± 0.790.0 ± 11.898.2 ± 2.285.0 ± 11.4

Downloads