imagining syntax

← efficiency_v1

ann_burst_hold14

done · peak unseen_mismatch 96.8

Referenced by

α schedule

α 3.0 ×150 α 1.0 ×150 α 2.1429 ×150 α 1.0 ×150 α 1.2857 ×150 α 1.0 ×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 α 1.4 ×150 α 1.4 ×150 α 1.4 ×150

direction
mixed
α range
1.0–3.0
stages
16

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 70.1 ± 16.162.2 ± 12.367.1 ± 15.660.3 ± 10.7
300 100.0 ± 0.040.7 ± 27.0100.0 ± 0.040.8 ± 27.3
450 99.8 ± 0.393.8 ± 15.199.5 ± 0.993.0 ± 14.8
600 100.0 ± 0.097.7 ± 3.1100.0 ± 0.096.8 ± 3.2
750 100.0 ± 0.097.2 ± 4.199.7 ± 0.594.6 ± 5.9
900 100.0 ± 0.098.1 ± 2.599.7 ± 0.994.6 ± 5.6
1050 99.9 ± 0.197.2 ± 2.498.7 ± 1.390.9 ± 6.0
1200 99.9 ± 0.196.0 ± 3.598.3 ± 1.486.2 ± 7.6
1350 99.8 ± 0.296.8 ± 2.798.0 ± 1.387.4 ± 6.9
1500 99.8 ± 0.296.6 ± 2.697.1 ± 1.781.2 ± 9.0
1650 99.9 ± 0.296.9 ± 2.396.9 ± 2.083.6 ± 10.0
1800 99.8 ± 0.295.7 ± 3.995.7 ± 2.780.2 ± 9.9
1950 99.9 ± 0.197.4 ± 2.095.3 ± 3.880.6 ± 9.4
2100 99.9 ± 0.295.2 ± 3.596.9 ± 2.475.5 ± 11.9
2250 99.8 ± 0.396.4 ± 3.394.7 ± 4.078.5 ± 12.4
2400 99.9 ± 0.196.6 ± 2.694.9 ± 2.378.4 ± 9.8

Downloads