s0_e2

done · peak unseen_mismatch 91.0

α schedule

α 0.0 ×50 α 0.087 ×50 α 0.1739 ×50 α 0.2609 ×50 α 0.3478 ×50 α 0.4348 ×50 α 0.5217 ×50 α 0.6087 ×50 α 0.6957 ×50 α 0.7826 ×50 α 0.8696 ×50 α 0.9565 ×50 α 1.0435 ×50 α 1.1304 ×50 α 1.2174 ×50 α 1.3043 ×50 α 1.3913 ×50 α 1.4783 ×50 α 1.5652 ×50 α 1.6522 ×50 α 1.7391 ×50 α 1.8261 ×50 α 1.913 ×50 α 2.0 ×50

direction: ascending
α range: 0.0–2.0
stages: 24

Figures

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 2.0

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

iteration	seen_match	seen_mismatch	unseen_match	unseen_mismatch
50	50.4 ± 1.6	49.9 ± 1.1	49.8 ± 1.9	49.6 ± 1.9
100	50.2 ± 1.2	50.0 ± 1.7	49.6 ± 1.2	49.4 ± 1.2
150	50.5 ± 3.5	49.0 ± 2.9	50.5 ± 3.5	48.8 ± 2.8
200	61.2 ± 18.6	43.8 ± 13.1	60.1 ± 19.2	44.0 ± 13.1
250	65.5 ± 21.2	40.3 ± 16.6	65.4 ± 21.2	40.2 ± 16.3
300	72.8 ± 23.2	38.1 ± 14.5	72.8 ± 23.0	40.0 ± 15.1
350	87.5 ± 19.8	29.9 ± 21.2	86.8 ± 20.6	29.8 ± 20.8
400	90.6 ± 18.8	32.1 ± 22.9	90.6 ± 18.8	33.3 ± 23.8
450	95.2 ± 14.3	31.6 ± 25.6	95.2 ± 14.3	32.3 ± 25.8
500	99.6 ± 1.2	25.7 ± 30.8	99.5 ± 1.6	25.6 ± 30.7
550	100.0 ± 0.0	29.3 ± 33.5	100.0 ± 0.0	29.1 ± 33.2
600	100.0 ± 0.0	28.8 ± 25.9	100.0 ± 0.0	28.9 ± 25.7
650	100.0 ± 0.0	29.8 ± 27.5	100.0 ± 0.0	30.0 ± 28.3
700	100.0 ± 0.0	26.2 ± 25.4	100.0 ± 0.0	26.7 ± 25.5
750	100.0 ± 0.0	35.1 ± 22.4	100.0 ± 0.0	35.4 ± 22.7
800	100.0 ± 0.0	43.5 ± 24.1	100.0 ± 0.0	43.4 ± 24.6
850	100.0 ± 0.0	46.2 ± 25.8	100.0 ± 0.0	46.6 ± 25.4
900	100.0 ± 0.0	59.1 ± 25.4	100.0 ± 0.0	58.6 ± 24.6
950	100.0 ± 0.0	71.3 ± 30.9	100.0 ± 0.0	70.8 ± 30.3
1000	100.0 ± 0.0	72.8 ± 26.8	100.0 ± 0.0	72.8 ± 26.1
1050	99.9 ± 0.2	82.1 ± 27.4	99.8 ± 0.4	81.2 ± 26.6
1100	99.7 ± 0.4	91.1 ± 14.8	99.0 ± 1.4	89.8 ± 14.8
1150	99.4 ± 0.8	92.8 ± 11.6	98.6 ± 1.9	91.0 ± 10.6
1200	98.9 ± 1.2	92.5 ± 11.6	97.6 ± 2.4	90.0 ± 11.2

s0_e2

α schedule

Figures

Results

Downloads