imagining syntax

← ascent_v1

s4_t2

done · peak unseen_mismatch 97.1

Referenced by

α schedule

α 4.0 ×19 α 4.0 ×19 α 4.0 ×19 α 4.0 ×19 α 4.0 ×19 α 4.0 ×19 α 4.0 ×19 α 4.0 ×19 α 4.0 ×19 α 4.0 ×19 α 4.0 ×19 α 4.0 ×19 α 1.0 ×19 α 1.0 ×19 α 1.0 ×19 α 1.0 ×19 α 1.0 ×19 α 1.0 ×19 α 1.0 ×19 α 1.0 ×19 α 1.0 ×19 α 1.0 ×19 α 1.0 ×19 α 1.0 ×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 1.0 ×18 α 1.0 ×18 α 1.0 ×18 α 1.0 ×18 α 1.0 ×18 α 1.0 ×18 α 1.0 ×18 α 1.0 ×18 α 1.0 ×18 α 1.0 ×18 α 1.0 ×18 α 1.0 ×18

direction
mixed
α range
1.0–4.0
stages
48

Figures

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

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 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 51.5 ± 1.649.7 ± 1.649.9 ± 2.049.7 ± 2.1
50 51.1 ± 1.749.9 ± 1.350.1 ± 1.550.4 ± 1.5
75 52.9 ± 2.150.9 ± 1.950.6 ± 1.750.0 ± 2.1
100 58.7 ± 5.555.5 ± 4.252.5 ± 4.051.1 ± 2.2
125 72.0 ± 10.167.7 ± 9.261.9 ± 9.357.9 ± 7.3
150 80.1 ± 7.877.2 ± 8.168.6 ± 9.464.3 ± 7.9
175 82.7 ± 4.580.6 ± 4.368.9 ± 8.666.1 ± 7.3
200 83.7 ± 3.781.9 ± 3.168.6 ± 7.166.1 ± 5.5
225 84.5 ± 3.883.1 ± 2.869.0 ± 6.566.3 ± 5.1
250 96.0 ± 3.395.2 ± 3.690.9 ± 7.988.9 ± 7.8
275 99.5 ± 0.498.9 ± 0.797.8 ± 1.696.6 ± 2.1
300 99.7 ± 0.399.3 ± 0.898.9 ± 0.897.1 ± 1.7
325 99.8 ± 0.299.4 ± 0.598.8 ± 1.097.0 ± 2.3
350 99.9 ± 0.299.4 ± 0.798.8 ± 1.496.4 ± 4.2
375 99.9 ± 0.299.5 ± 0.599.0 ± 1.096.2 ± 3.6
400 100.0 ± 0.199.5 ± 0.499.4 ± 0.896.8 ± 2.9
450 100.0 ± 0.199.6 ± 0.399.2 ± 0.796.1 ± 2.6
500 99.7 ± 0.298.9 ± 0.898.1 ± 1.193.5 ± 3.5
550 99.8 ± 0.299.2 ± 0.597.3 ± 1.592.7 ± 3.4
600 99.8 ± 0.299.0 ± 0.797.8 ± 2.291.3 ± 5.4
650 99.8 ± 0.198.8 ± 0.997.3 ± 2.089.4 ± 7.4
700 100.0 ± 0.099.4 ± 0.699.6 ± 0.494.5 ± 4.4
750 100.0 ± 0.198.9 ± 1.099.0 ± 0.991.4 ± 6.7
800 100.0 ± 0.199.0 ± 1.099.3 ± 0.691.1 ± 7.5
850 100.0 ± 0.098.6 ± 2.299.0 ± 1.490.2 ± 9.4
900 99.9 ± 0.198.2 ± 2.198.4 ± 1.286.3 ± 9.6

Downloads