imagining syntax

← schedopt_v5_confirm

free-6217439f

done · peak unseen_mismatch 92.6

α schedule

α 5.0 ×10 α 4.1 ×10 α 4.3 ×10 α 4.9 ×10 α 4.5 ×10 α 4.4 ×10 α 4.0 ×10 α 4.5 ×10 α 4.7 ×10 α 5.0 ×20 α 4.3 ×10 α 4.6 ×10 α 4.3 ×10 α 4.8 ×10 α 4.4 ×10 α 4.2 ×10 α 1.2 ×10 α 0.0 ×10 α 0.6 ×10 α 0.0 ×10 α 0.6 ×10 α 2.1 ×10 α 3.8 ×20 α 5.0 ×10 α 1.7 ×10 α 1.3 ×10 α 0.0 ×120

direction
mixed
α range
0.0–5.0
stages
27

Figures

schedule.png
schedule.png

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 0.0

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
100 60.5 ± 7.856.7 ± 6.254.2 ± 5.951.8 ± 3.6
110 65.4 ± 9.261.6 ± 7.956.9 ± 7.553.6 ± 5.3
120 70.0 ± 10.266.4 ± 9.159.7 ± 8.856.2 ± 6.8
130 74.1 ± 9.670.6 ± 8.662.0 ± 9.058.4 ± 7.1
140 76.9 ± 7.973.7 ± 7.463.4 ± 8.960.1 ± 6.9
150 78.6 ± 6.676.1 ± 5.964.1 ± 8.261.1 ± 6.7
160 80.0 ± 5.677.5 ± 5.064.4 ± 7.761.5 ± 6.4
170 80.6 ± 4.778.3 ± 4.364.6 ± 7.062.1 ± 6.0
180 84.6 ± 5.682.3 ± 5.370.5 ± 8.867.4 ± 8.0
190 90.3 ± 7.287.8 ± 7.582.4 ± 10.878.7 ± 10.7
200 95.9 ± 6.493.8 ± 7.292.9 ± 8.989.8 ± 9.5
210 97.3 ± 6.093.6 ± 9.595.9 ± 7.591.0 ± 10.8
220 97.0 ± 4.983.0 ± 17.796.1 ± 6.380.1 ± 17.8
230 98.7 ± 2.779.4 ± 21.497.9 ± 3.376.7 ± 21.3
240 98.7 ± 1.987.7 ± 12.797.8 ± 2.484.0 ± 12.2
250 96.5 ± 2.193.4 ± 1.992.5 ± 4.387.3 ± 3.3
260 93.1 ± 1.390.5 ± 0.388.0 ± 3.781.5 ± 2.7
270 96.4 ± 0.892.6 ± 1.092.7 ± 3.787.4 ± 3.3
280 98.6 ± 0.195.0 ± 0.695.5 ± 0.788.5 ± 1.1
290 99.2 ± 0.295.0 ± 2.198.4 ± 1.090.3 ± 1.6
300 100.0 ± 0.088.7 ± 10.6100.0 ± 0.085.3 ± 13.6
310 100.0 ± 0.095.4 ± 0.0100.0 ± 0.092.6 ± 0.0

Downloads