imagining syntax

← slopesweep_v1

s1_e1

done · peak unseen_mismatch 89.5

α schedule

α 1.0 ×50 α 1.0 ×50 α 1.0 ×50 α 1.0 ×50 α 1.0 ×50 α 1.0 ×50 α 1.0 ×50 α 1.0 ×50 α 1.0 ×50 α 1.0 ×50 α 1.0 ×50 α 1.0 ×50 α 1.0 ×50 α 1.0 ×50 α 1.0 ×50 α 1.0 ×50 α 1.0 ×50 α 1.0 ×50 α 1.0 ×50 α 1.0 ×50 α 1.0 ×50 α 1.0 ×50 α 1.0 ×50 α 1.0 ×50

direction
constant
α range
1.0–1.0
stages
24

Figures

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

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 1.0

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
50 50.2 ± 1.049.1 ± 1.549.8 ± 1.651.2 ± 1.6
100 50.8 ± 2.649.7 ± 1.751.3 ± 1.948.7 ± 2.4
150 56.6 ± 14.648.7 ± 1.455.0 ± 15.050.6 ± 1.6
200 71.1 ± 19.739.3 ± 10.769.6 ± 19.940.4 ± 11.2
250 90.7 ± 18.527.3 ± 16.490.4 ± 19.026.9 ± 15.4
300 99.6 ± 0.728.0 ± 20.899.5 ± 0.828.2 ± 20.6
350 100.0 ± 0.017.0 ± 17.0100.0 ± 0.017.1 ± 17.7
400 100.0 ± 0.021.9 ± 21.9100.0 ± 0.022.3 ± 22.6
450 100.0 ± 0.029.9 ± 17.6100.0 ± 0.030.5 ± 18.3
500 100.0 ± 0.027.9 ± 17.2100.0 ± 0.027.7 ± 17.6
550 100.0 ± 0.046.4 ± 18.8100.0 ± 0.047.0 ± 19.0
600 100.0 ± 0.040.6 ± 27.2100.0 ± 0.041.1 ± 26.8
650 100.0 ± 0.057.2 ± 27.0100.0 ± 0.056.8 ± 26.8
700 100.0 ± 0.041.5 ± 22.3100.0 ± 0.041.9 ± 21.7
750 100.0 ± 0.050.7 ± 28.9100.0 ± 0.051.0 ± 28.9
800 100.0 ± 0.057.3 ± 26.0100.0 ± 0.056.8 ± 26.3
850 100.0 ± 0.068.6 ± 20.2100.0 ± 0.069.3 ± 20.5
900 100.0 ± 0.068.1 ± 25.9100.0 ± 0.067.9 ± 26.5
950 100.0 ± 0.070.5 ± 23.2100.0 ± 0.070.6 ± 23.4
1000 100.0 ± 0.076.0 ± 19.5100.0 ± 0.076.0 ± 19.3
1050 100.0 ± 0.076.3 ± 28.6100.0 ± 0.076.0 ± 28.9
1100 100.0 ± 0.084.8 ± 19.5100.0 ± 0.084.2 ± 19.6
1150 100.0 ± 0.090.5 ± 15.5100.0 ± 0.089.5 ± 16.4
1200 100.0 ± 0.087.8 ± 24.9100.0 ± 0.086.9 ± 25.0

Downloads