Adaptive Online Sequential ELM for Concept Drift Tackling

Table 7

Predictive accuracy performance in AOS-ELM for MNIST using different parameters and before and after increase. Each experiment is repeated 100x trials to get the probability of predictive accuracy ≤ 50%.

Scenario			Batch size		Before	After	Pred. acc. ≤ 50%
RD	630	500	1000	10	630	1130	0%
	630	500	500	10	1130	1122	7%
	630	500	100	10	1130	614	5%
	630	500	10	10	640	640	0%
	630	100	1000	10	630	730	0%
	630	100	500	10	730	730	0%
	630	100	100	10	730	730	0%
	630	100	10	10	640	640	0%
	2000	1000	100	5	2000	1868	3%
	2000	1000	100	1	2000	1947	16%
	2000	1000	100	0.5	2000	1946	17%
	2000	1000	100	0.05	2000	2100	0%
VD	666	500	500	5	666	1166	0%
	666	500	100	5	666	1166	0%
	666	100	500	5	666	766	0%
	666	100	100	5	666	766	0%
	666	50	500	5	666	716	0%
	666	50	100	5	666	716	0%