imagining syntax

← schedopt_v5_confirm

free-241ff30e

done · peak unseen_mismatch 93.4

α schedule

α 4.9 ×10 α 4.3 ×10 α 5.0 ×10 α 4.7 ×10 α 5.0 ×10 α 4.8 ×10 α 3.6 ×10 α 4.4 ×10 α 5.0 ×10 α 4.3 ×10 α 4.8 ×10 α 4.0 ×10 α 4.7 ×10 α 4.0 ×10 α 5.0 ×20 α 4.2 ×10 α 1.0 ×10 α 0.0 ×10 α 0.4 ×10 α 0.0 ×20 α 2.5 ×10 α 3.3 ×10 α 4.0 ×10 α 5.0 ×10 α 3.0 ×10 α 0.4 ×10 α 0.0 ×10 α 0.1 ×110

direction
mixed
α range
0.0–5.0
stages
28

Figures

schedule.png
schedule.png

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 0.1

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
100 61.3 ± 7.857.6 ± 6.454.6 ± 5.851.8 ± 3.5
110 66.5 ± 9.062.9 ± 7.957.0 ± 7.253.9 ± 5.0
120 71.4 ± 9.367.7 ± 8.759.9 ± 8.556.6 ± 6.4
130 75.6 ± 8.572.1 ± 8.062.8 ± 8.659.5 ± 7.0
140 78.1 ± 7.375.2 ± 6.764.1 ± 8.360.7 ± 6.8
150 79.4 ± 6.276.7 ± 5.764.3 ± 8.161.5 ± 6.8
160 79.9 ± 4.977.5 ± 4.263.8 ± 7.061.4 ± 6.1
170 80.6 ± 4.078.5 ± 3.464.1 ± 6.461.8 ± 5.5
180 84.5 ± 5.082.2 ± 4.670.1 ± 8.267.3 ± 7.4
190 90.2 ± 6.687.7 ± 6.882.2 ± 10.779.0 ± 10.4
200 96.5 ± 3.994.0 ± 5.993.8 ± 6.290.3 ± 7.8
210 99.1 ± 1.194.6 ± 8.998.4 ± 2.092.6 ± 9.8
220 99.7 ± 0.777.5 ± 14.599.7 ± 0.875.5 ± 14.2
230 99.9 ± 0.285.4 ± 14.199.9 ± 0.384.0 ± 13.5
240 99.9 ± 0.289.8 ± 10.699.7 ± 0.488.0 ± 10.8
250 99.9 ± 0.192.5 ± 10.899.7 ± 0.390.6 ± 11.2
260 100.0 ± 0.093.0 ± 0.099.9 ± 0.090.0 ± 0.0
270 100.0 ± 0.097.4 ± 0.099.8 ± 0.093.4 ± 0.0

Downloads