Evaluation of maxout activations in deep learning across several big data domains

Castaneda, Gabriel; Morris, Paul; Khoshgoftaar, Taghi M.

doi:10.1186/s40537-019-0233-0

Journal of Big Data

Table 1 CNN and NN Configurations

From: Evaluation of maxout activations in deep learning across several big data domains

Layer	A Medicare Part B Medicare Part D DMEPOS Combined CMS	B MNIST FMNIST CIFAR10 CIFAR100	C LFW	D Sent140 Amazon Yelp	E GSC IRMAS IDMT-SMT-Audio-Effects	F MS-Celeb
Input	123 123 123 123	28 × 28 × 1 28 × 28 × 1 32 × 32 × 3 32 × 32 × 3	128 × 128 × 1	8 × 140 × 1 8 × 500 × 1 8 × 500 × 1	129 × 71 × 1 129 × 71 × 1 100 × 100 × 1	128 × 128 × 1
Conv	N/A	f = 10 k = [5,5] p = none	f = 32 k = [7,7]	f = 64 k = [3,3]	f = 32 k = [5,5]	f = 48 k = [5,5]
Pool	N/A	k = [2,2] s = [2,2]	k = [2,2] s = [2,2]	N/A	k = [2,2] s = [2,2]	k = [2,2] s = [2,2]
C-NIN	N/A	N/A	N/A	N/A	N/A	f = 48 k = [1,1]
Conv	N/A	f = 20 k = [5, 5] p = none	f = 64 k = [5, 5]	f = 64 k = [3,3]	f = 64 k = [3,3]	f = 96 k = [3,3]
Pool	N/A	N/A	k = [2,2] s = [2,2]	N/A	N/A	k = [2,2] s = [2,2]
C-NIN	N/A	N/A	N/A	N/A	N/A	f = 96 k = [1,1]
Conv	N/A	N/A	f = 64 k = [3,3]	f = 64 k = [3,3]	f = 64 k = [3,3]	f = 64 k = [3,3]
Pool	N/A	N/A	N/A	k = [2,2] s = [2,2]	k = [2,2] s = [2,2]	k = [2,2] s = [2,2]
C-NIN	N/A	N/A	N/A	N/A	N/A	f = 192 k = [1,1]
Conv	N/A	N/A	N/A	f = 64 k = [3,3]	f = 64 k = [3,3]	f = 128 k = [3,3]
C-NIN	N/A	N/A	N/A	N/A	N/A	f = 128 k = [1,1]
Conv	N/A	N/A	N/A	f = 64 k = [3,3]	f = 64 k = [3,3]	f = 64 k = [3,3]
Conv	N/A	N/A	N/A	f = 64 k = [3,3]	N/A	N/A
Pool	N/A	N/A	N/A	k = [2,2] s = [2,2]	k = [2,2] s = [2,2]	k = [2,2] s = [2,2]
FC	n = 512	N/A	n = 256	n = 512	n = 512	n = 256
DO	kp = 0.5	kp = 0.5	kp = 0.5	kp = 0.5	kp = 0.5	kp = 0.5
Pool	N/A	k = [2,2] s = [2,2]	k = [2,2] s = [2,2]	N/A	N/A	N/A
FC	n = 64	n = 50 n = 250^a	N/A	n = 512	n = 512	n = 256
DO	kp = 0.5	kp = 0.5	kp = 0.5	kp = 0.5	kp = 0.5	kp = 0.5
FC	n = 2	n = 10 n = 100^a	n = 2	n = 2	n = 35 n = 11^b	n = 1000

Convolutional layers (Conv) indicate the number of filters (f=), the kernel size (k=) and the padding (p=). Convolution network in network layers (C-NIN) indicate the number of filters (f=), and the kernel size (k=). Max-pool layers (Pool) indicate the kernel size (k=) and the stride (s=). Dropout layers (DO) show the applied keep probability (kp=), and the fully-connected layers (FC) display the number of neurons (n=)
^aNeurons applied only to the CIFAR-100 dataset
^bNeurons applied only to the IRMAS and IDMT-SMT-Audio-Effects datasets

Back to article page