imagining syntax

← refine_v1

bursty_K24

done · peak unseen_mismatch 96.3

α schedule

α 3.0 ×38 α 3.0 ×38 α 3.0 ×38 α 3.0 ×38 α 1.0 ×38 α 1.0 ×38 α 1.0 ×38 α 1.0 ×38 α 2.1429 ×38 α 2.1429 ×38 α 2.1429 ×38 α 2.1429 ×38 α 1.0 ×37 α 1.0 ×37 α 1.0 ×37 α 1.0 ×37 α 1.2857 ×37 α 1.2857 ×37 α 1.2857 ×37 α 1.2857 ×37 α 1.0 ×37 α 1.0 ×37 α 1.0 ×37 α 1.0 ×37

direction
mixed
α range
1.0–3.0
stages
24

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 51.8 ± 2.549.6 ± 1.850.3 ± 2.049.9 ± 2.2
75 52.1 ± 3.649.8 ± 2.352.1 ± 3.250.5 ± 1.5
100 56.2 ± 9.250.8 ± 2.253.7 ± 9.050.8 ± 1.8
125 64.6 ± 15.457.6 ± 10.061.8 ± 15.255.1 ± 8.7
150 78.6 ± 15.869.5 ± 13.974.3 ± 17.565.3 ± 13.8
175 90.0 ± 15.674.9 ± 15.588.9 ± 16.272.2 ± 15.8
200 99.1 ± 3.570.3 ± 19.599.2 ± 2.969.8 ± 19.1
225 100.0 ± 0.068.9 ± 19.8100.0 ± 0.068.5 ± 19.7
250 100.0 ± 0.064.1 ± 23.9100.0 ± 0.063.2 ± 23.7
275 100.0 ± 0.065.7 ± 33.0100.0 ± 0.065.0 ± 32.6
300 100.0 ± 0.068.6 ± 30.7100.0 ± 0.068.4 ± 30.3
325 100.0 ± 0.076.1 ± 28.9100.0 ± 0.275.9 ± 28.8
350 99.9 ± 0.286.9 ± 23.299.6 ± 0.885.5 ± 23.7
375 99.8 ± 0.592.1 ± 15.399.0 ± 1.890.5 ± 15.3
400 99.4 ± 0.896.4 ± 4.798.0 ± 2.692.8 ± 7.1
450 99.1 ± 1.197.5 ± 1.997.5 ± 3.193.7 ± 5.5
500 100.0 ± 0.096.8 ± 12.799.9 ± 0.296.3 ± 12.1
550 100.0 ± 0.095.0 ± 14.399.9 ± 0.293.5 ± 14.5
600 100.0 ± 0.194.0 ± 12.099.9 ± 0.290.9 ± 13.0
650 100.0 ± 0.195.3 ± 10.799.5 ± 0.891.1 ± 11.9
700 99.9 ± 0.197.6 ± 2.799.2 ± 1.292.7 ± 8.7
750 99.9 ± 0.297.7 ± 3.198.5 ± 1.690.6 ± 11.3
800 100.0 ± 0.096.5 ± 8.399.8 ± 0.591.4 ± 12.9
850 100.0 ± 0.096.8 ± 4.299.7 ± 0.689.8 ± 13.8
900 100.0 ± 0.195.6 ± 7.799.6 ± 0.786.8 ± 17.3

Downloads