Table 2.

NAS search space networks performance and statistical features by dataset.

Dataset	Maximum performance	Minimum performance	Average performance	SD
MNIST	98.69%1	8.92%2	97.84%3	0.65%4
CIFAR-10	56.31%5	10.00%6	40.44%	12.40%

¹The best performing neural network architecture against this dataset consisted of 4 layers, 768 neurons per layer, using the ELU activation function, and the RMSPROP optimizer.

²The worst performing neural network architecture against this dataset consisted of 3 layers, 1,024 neurons per layer, using the RELU activation function, and the ADAGRAD optimizer.

³This statistical value was computed after discarding 12 outlier performance data points (1.79% of the total data collected). If this discarded data is included, the average performance becomes 96.27%.

⁴This statistical value was computed after discarding 12 outlier performance data points (1.79% of the total data collected). If this discarded data is included, the standard deviation becomes 11.67%.

⁵The best performing neural network architecture against this dataset consisted of 2 layers, 512 neurons per layer, using the ELU activation function, and the ADAMAX optimizer.

⁶The worst performing neural network architecture against this dataset consisted of 4 layers, 1,024 neurons per layer, using the SIGMOID activation function, and the SGD optimizer.