Table 2.

Empirical coverage of the human activity recognition (HAR) using smartphones and the Massachusetts Institute of Technology-Beth Israel Hospital (MIT-BIH) arrhythmia data sets by different normalization methods. Target coverage was set before training.

Target coverage	HAR using smartphones data set				MIT-BIH arrhythmia data set
	Normalization method of selective prediction				Normalization method of selective prediction
	UBSa	BNb	Without normalizationc	UBS		BN	Without normalization
0.95, mean (SD)	0.9660 (0.0029)	0.9996 (0.0001)	0.9986 (0.0002)	0.9564 (0.0019)		0.9680 (0.0067)	1.0000 (0)
0.90, mean (SD)	0.9053 (0.0035)	0.9980 (0.0001)	0.9984 (0.0001)	0.9084 (0.0055)		0.9998 (0.0001)	1.0000 (0)
0.85, mean (SD)	0.8582 (0.0007)	0.9237 (0.0026)	0.9986 (0.0002)	0.8888 (0.0016)		0.9518 (0.0001)	1.0000 (0)
Average violation, %	0.98	7.38	9.85	1.79		7.32	10.00

^aUBS: unit-wise batch standardization.

^bBN: batch normalization (a normalization method using the mean and variance obtained from the input batch).

^cWithout normalization means that there was no normalization in the selection function structure.