. 2018 Jun 19;9:2383. doi: 10.1038/s41467-018-04316-3

Table 3.

Summarization of the experiments with MLP variants

Dataset	Data augmentation	Architecture	Activation	MLP		MLP_FixProb		SET-MLP
Dataset	Data augmentation	Architecture	Activation	Accuracy [%]	n ^W	Accuracy [%]	n ^W	Accuracy [%]	n ^W
MNIST	No	784-1000-1000-1000-10	SReLU	98.55	2,794,000	97.68	89,797	98.74	89,797
CIFAR10	Yes	3072-4000-1000-4000-10	SReLU	68.70	20,328,000	62.19	278,630	74.84	278,630
HIGGS	No	28-1000-1000-1000-2	SReLU	78.44	2,038,000	76.69	80,614	78.47	80,614

On each dataset, we report the best classification accuracy obtained by each model on the test data. n^W represents the number of weights in the model. The only difference between the three models is the network topology, i.e. MLP has fully connected layers, MLP_FixProb has sparse layers with Erdös–Rényi fixed topology, and SET-MLP has sparse evolutionary layers trained with SET