. 2020 Nov 9;21:771. doi: 10.1186/s12864-020-07181-x

Table 2.

The best deep neural network architecture selected based on prediction correlation on the tuning set for each sub-sampling of the training set

Size (%)	Deep neural network architecture
Size (%)	Number of layers	Number of units per layer^a	L2^b	Dropout rate^c	Accuracy	MSEP^d
1	4	5000⁽¹⁾-1⁽²⁾-600⁽³⁾-800⁽⁴⁾	0.0600	1.0	0.090	30,589.3
3	4	5000⁽¹⁾-300⁽²⁾-200⁽³⁾-4000⁽⁴⁾	0.0675	1.0	0.137	29,649.9
5	3	400⁽¹⁾-200⁽²⁾ -900⁽³⁾	0.0100	0.5	0.145	30,408.7
7	2	500⁽¹⁾-2000⁽²⁾	0.0450	0.8	0.166	29,062.4
10	2	800⁽¹⁾-100⁽²⁾	0.0025	0.6	0.200	28,440.9
15	2	800⁽¹⁾-900⁽²⁾	0.0050	0.5	0.236	27,755.0
20	4	600⁽¹⁾-100⁽²⁾-500⁽³⁾-700⁽⁴⁾	0.0325	0.5	0.226	28,849.5
30	1	1000⁽¹⁾	0.0100	0.7	0.274	27,025.5
40	1	2000⁽¹⁾	0.0800	0.6	0.285	26,877.4
50	3	600⁽¹⁾-4000⁽²⁾ -100⁽³⁾	0.0975	0.5	0.285	27,250.3
60	1	300⁽¹⁾	0.0800	0.8	0.304	26,622.3
70	1	400⁽¹⁾	0.0800	0.5	0.309	26,506.4
80	1	800⁽¹⁾	0.0925	0.7	0.308	26,484.5
90	1	400⁽¹⁾	0.0800	0.5	0.307	26,710.1
100	1	500⁽¹⁾	0.0600	1.0	0.322	26,264.8

^aThe number in parenthesis represents the corresponding hidden layer

^bL2 = ridge regularization

^cDropout rate was applied in all layers, except for the output layer

^dMSEP = mean square error of prediction