imagining syntax

← refine_v1

bursty_K12

done · peak unseen_mismatch 90.9

Referenced by

α schedule

α 3.0 ×75 α 3.0 ×75 α 1.0 ×75 α 1.0 ×75 α 2.1429 ×75 α 2.1429 ×75 α 1.0 ×75 α 1.0 ×75 α 1.2857 ×75 α 1.2857 ×75 α 1.0 ×75 α 1.0 ×75

direction
mixed
α range
1.0–3.0
stages
12

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
25 51.1 ± 1.249.8 ± 2.049.8 ± 1.350.4 ± 1.4
50 51.5 ± 1.850.0 ± 1.750.5 ± 1.149.7 ± 1.5
75 52.2 ± 3.349.5 ± 1.650.9 ± 2.749.6 ± 2.7
100 55.3 ± 8.351.2 ± 3.053.7 ± 7.950.4 ± 2.4
125 65.3 ± 13.057.9 ± 10.662.4 ± 12.854.7 ± 8.4
150 80.2 ± 15.368.3 ± 13.976.6 ± 15.964.1 ± 12.4
175 93.6 ± 10.869.0 ± 16.992.8 ± 11.668.0 ± 15.7
200 98.7 ± 4.868.6 ± 20.998.6 ± 5.567.7 ± 20.5
225 100.0 ± 0.263.7 ± 26.299.9 ± 0.263.5 ± 26.1
250 100.0 ± 0.064.1 ± 30.4100.0 ± 0.064.8 ± 29.7
275 100.0 ± 0.064.4 ± 34.2100.0 ± 0.064.2 ± 33.9
300 100.0 ± 0.061.4 ± 34.9100.0 ± 0.061.7 ± 34.5
325 100.0 ± 0.065.6 ± 33.099.9 ± 0.366.2 ± 32.3
350 99.8 ± 0.472.2 ± 28.899.8 ± 0.572.0 ± 27.8
375 99.8 ± 0.384.2 ± 26.399.6 ± 0.882.8 ± 25.9
400 99.5 ± 0.985.0 ± 24.198.4 ± 2.883.3 ± 23.4
450 99.4 ± 0.889.2 ± 20.497.7 ± 3.086.2 ± 19.5
500 100.0 ± 0.091.4 ± 19.399.9 ± 0.290.9 ± 19.4
550 100.0 ± 0.188.8 ± 24.099.8 ± 0.487.2 ± 23.3
600 100.0 ± 0.086.7 ± 22.499.9 ± 0.385.4 ± 21.6
650 99.9 ± 0.190.2 ± 23.299.1 ± 1.586.3 ± 22.5
700 99.9 ± 0.194.7 ± 11.899.0 ± 1.390.1 ± 11.9
750 100.0 ± 0.194.9 ± 11.899.2 ± 1.490.6 ± 13.5
800 100.0 ± 0.090.6 ± 20.799.6 ± 1.284.8 ± 22.0
850 100.0 ± 0.087.7 ± 23.199.6 ± 0.981.5 ± 24.1
900 100.0 ± 0.091.9 ± 14.499.7 ± 0.785.1 ± 19.2

Downloads