Table 4.

Comparison of training and test accuracy for various methods.

Principal components	Training accuracy (%, mean ± std)	Test accuracy (%, mean ± std)
All [same as Finn et al. (2015)]	88.65 ± 1.76	88.98 ± 1.72
2:end	94.30 ± 1.35	71.76 ± 8.76
11:end	96.74 ± 1.00	69.61 ± 8.94
21:end	95.03 ± 1.90	69.44 ± 8.99
31:end	71.97 ± 6.08	68.95 ± 9.07
41:end	72.77 ± 1.74	65.70 ± 9.59
Principal features subspace	96.23 ± 2.24	93.11 ± 3.61

Numbers in the first row use the entire matrix; the next five rows use a subset of the principal components; the last row corresponds to our method. Our method achieves almost optimal training set accuracy and significantly better test set accuracy compared to competing methods.