Table A2.

pCNN architecture and parameter size.

No	Type	Input size	Patch size/stride	Padding	Output size	Param. mem.
1	Convolution 1	224 × 224 × 2	11 × 11/4	1	54 × 54 × 64	61 KB
2	ReLU 1	54 × 54 × 64			54 × 54 × 64
3	Normalization 1	54 × 54 × 64			54 × 54 × 64
4	Mean pool 1	54 × 54 × 64	2 × 2/2	0	27 × 27 × 64
5	Convolution 2	27 × 27 × 64	5 × 5/1	2	27 × 27 × 64	400 KB
6	ReLU 2	27 × 27 × 64			27 × 27 × 64
7	Normalization 2	27 × 27 × 64			54 × 54 × 64
8	Mean pool 2	27 × 27 × 64	2 × 2/2	0	13 × 13 × 64
9	Convolution 3	13 × 13 × 64	3 × 3/1	1	13 × 13 × 64	144 KB
10	ReLU 3	13 × 13 × 64			13 × 13 × 64
11	Convolution 4	13 × 13 × 64	3 × 3/1	1	13 × 13 × 64	144 KB
12	ReLU 4	13 × 13 × 64			13 × 13 × 64
13	Convolution 5	13 × 13 × 64	3 × 3/1	1	13 × 13 × 64	144 KB
14	ReLU 5	13 × 13 × 64			162 × 1 × 36
15	Mean pool	13 × 13 × 64	2 × 2/2		6 × 6 × 64
16	Fully connected layer 6	6 × 6 × 64	6 × 6/1	0	1 × 1 × 100	900 KB
17	ReLU 6	1 × 1 × 100			1 × 1 × 100
18	Fully connected layer 7	1 × 1 × 100	1 × 1/1		1 × 1 × 2	408 B
19	Softmax classifier	1 × 1 × 2
Total						1.79 MB