imagining syntax

← ascent_v1

s5_v07_t2

done · peak unseen_mismatch 99.8

Referenced by

α schedule

α 5.0 ×19 α 5.0 ×19 α 5.0 ×19 α 5.0 ×19 α 5.0 ×19 α 5.0 ×19 α 5.0 ×19 α 5.0 ×19 α 5.0 ×19 α 5.0 ×19 α 5.0 ×19 α 5.0 ×19 α 0.7 ×19 α 0.7 ×19 α 0.7 ×19 α 0.7 ×19 α 0.7 ×19 α 0.7 ×19 α 0.7 ×19 α 0.7 ×19 α 0.7 ×19 α 0.7 ×19 α 0.7 ×19 α 0.7 ×19 α 1.2857 ×19 α 1.2857 ×19 α 1.2857 ×19 α 1.2857 ×19 α 1.2857 ×19 α 1.2857 ×19 α 1.2857 ×19 α 1.2857 ×19 α 1.2857 ×19 α 1.2857 ×19 α 1.2857 ×19 α 1.2857 ×19 α 0.7 ×18 α 0.7 ×18 α 0.7 ×18 α 0.7 ×18 α 0.7 ×18 α 0.7 ×18 α 0.7 ×18 α 0.7 ×18 α 0.7 ×18 α 0.7 ×18 α 0.7 ×18 α 0.7 ×18

direction
mixed
α range
0.7–5.0
stages
48

Figures

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

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 0.7

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
25 51.5 ± 1.549.9 ± 1.349.8 ± 1.949.6 ± 2.2
50 50.9 ± 1.850.0 ± 1.250.2 ± 1.650.4 ± 1.5
75 54.0 ± 1.951.6 ± 1.951.1 ± 1.750.1 ± 2.2
100 62.4 ± 6.258.5 ± 4.654.2 ± 4.351.8 ± 2.1
125 74.4 ± 8.770.9 ± 8.462.6 ± 8.959.0 ± 7.4
150 78.4 ± 4.576.4 ± 4.064.7 ± 6.361.9 ± 5.5
175 79.9 ± 3.978.6 ± 3.164.7 ± 5.862.6 ± 5.3
200 80.1 ± 3.378.6 ± 2.863.1 ± 4.861.8 ± 3.8
225 80.3 ± 3.378.9 ± 2.463.8 ± 4.362.7 ± 4.1
250 92.0 ± 5.990.8 ± 6.084.2 ± 11.382.7 ± 10.7
275 99.7 ± 0.499.5 ± 0.699.3 ± 0.898.6 ± 1.3
300 100.0 ± 0.199.8 ± 0.299.9 ± 0.299.5 ± 0.7
325 100.0 ± 0.199.8 ± 0.399.8 ± 0.499.5 ± 0.8
350 100.0 ± 0.099.9 ± 0.5100.0 ± 0.199.4 ± 1.1
375 100.0 ± 0.0100.0 ± 0.1100.0 ± 0.299.7 ± 0.4
400 100.0 ± 0.0100.0 ± 0.099.9 ± 0.299.8 ± 0.4
450 100.0 ± 0.099.9 ± 0.299.9 ± 0.399.4 ± 0.6
500 99.8 ± 0.299.0 ± 0.598.5 ± 1.195.2 ± 2.2
550 99.8 ± 0.299.2 ± 0.497.6 ± 1.193.7 ± 2.0
600 99.8 ± 0.199.2 ± 0.698.2 ± 1.393.9 ± 3.7
650 99.8 ± 0.299.2 ± 0.697.1 ± 1.691.9 ± 4.3
700 100.0 ± 0.099.8 ± 0.5100.0 ± 0.198.5 ± 2.6
750 100.0 ± 0.099.4 ± 1.099.9 ± 0.195.5 ± 4.3
800 100.0 ± 0.097.8 ± 3.499.8 ± 0.391.9 ± 8.1
850 100.0 ± 0.094.0 ± 10.899.8 ± 0.485.5 ± 16.0
900 100.0 ± 0.095.1 ± 5.899.3 ± 0.885.5 ± 11.1

Downloads