imagining syntax

← schedopt_v5_confirm

free-e059ed86

done · peak unseen_mismatch 94.8

α schedule

α 4.8 ×10 α 4.3 ×10 α 5.0 ×10 α 4.0 ×10 α 4.3 ×10 α 4.2 ×10 α 4.8 ×10 α 4.0 ×10 α 4.6 ×10 α 4.0 ×10 α 4.7 ×10 α 4.4 ×10 α 4.1 ×10 α 4.8 ×10 α 5.0 ×10 α 4.5 ×10 α 4.4 ×10 α 1.1 ×10 α 0.8 ×10 α 0.3 ×10 α 0.4 ×10 α 0.9 ×10 α 1.9 ×10 α 5.0 ×10 α 4.2 ×10 α 4.5 ×10 α 2.1 ×10 α 1.4 ×10 α 0.0 ×10 α 0.4 ×110

direction
mixed
α range
0.0–5.0
stages
30

Figures

schedule.png
schedule.png

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 0.4

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
100 59.9 ± 7.556.5 ± 6.154.1 ± 5.651.6 ± 3.4
110 65.0 ± 9.061.1 ± 7.956.4 ± 7.653.6 ± 5.0
120 70.0 ± 9.666.2 ± 8.559.5 ± 8.756.2 ± 6.5
130 74.3 ± 9.570.8 ± 8.662.2 ± 9.059.0 ± 7.0
140 77.3 ± 8.474.4 ± 7.664.2 ± 8.960.8 ± 7.1
150 79.0 ± 6.876.4 ± 6.064.6 ± 8.261.6 ± 6.7
160 80.0 ± 5.577.6 ± 5.064.4 ± 7.161.7 ± 6.1
170 80.8 ± 4.678.4 ± 4.064.9 ± 6.862.2 ± 5.8
180 84.7 ± 5.682.2 ± 5.270.9 ± 8.867.8 ± 7.9
190 91.5 ± 5.989.4 ± 6.283.0 ± 9.979.6 ± 9.9
200 96.3 ± 4.993.9 ± 6.293.1 ± 7.189.7 ± 8.3
210 98.1 ± 3.092.9 ± 9.497.0 ± 4.290.3 ± 10.4
220 99.5 ± 0.786.5 ± 13.999.0 ± 1.883.2 ± 14.4
230 99.8 ± 0.484.9 ± 12.799.6 ± 0.482.1 ± 12.5
240 99.8 ± 0.286.5 ± 8.999.5 ± 0.882.4 ± 9.6
250 97.8 ± 2.686.7 ± 10.095.5 ± 5.179.0 ± 11.7
260 97.5 ± 1.490.1 ± 5.293.8 ± 3.181.7 ± 8.0
270 98.2 ± 1.292.1 ± 3.595.3 ± 2.183.8 ± 8.3
280 99.2 ± 0.495.7 ± 2.498.5 ± 1.091.9 ± 4.4
290 99.9 ± 0.196.7 ± 1.9100.0 ± 0.094.8 ± 2.8

Downloads