Table A1.

gCNN architecture and parameter size.

No.	Type	Input size	Patch size/stride	Output size	Param. mem.
1	Convolution 1	40962 × 25 × 2	1 × 25/1	40962 × 1 × 36	7 KB
2	Batch normalization 1			40962 × 1 × 36	576 B
3	ReLU 1			40962 × 1 × 36
4	Mean pool 1	40962 × 1 × 36	1 × 6 or* 1 × 7/1	10242 × 1 × 36
5	Convolution 2	10242 × 25 × 36	1 × 25/1	10242 × 1 × 36	127 KB
6	Batch normalization 2			10242 × 1 × 36	576 B
7	ReLU 2			10242 × 1 × 36
8	Mean pool 2	10242 × 1 × 36	1 × 6 or 1 × 7/1	2562 × 1 × 36
9	Convolution 3	2562 × 25 × 36	1 × 25/1	2562 × 1 × 36	127 KB
10	Batch normalization 3			2562 × 1 × 36	576 B
11	ReLU 3			2562 × 1 × 36
12	Mean pool 3	2562 × 1 × 36	1 × 6 or 1 × 7/1	642 × 1 × 36
13	Convolution 4	642 × 25 × 36	1 × 25/1	642 × 1 × 36	127 KB
14	Batch normalization 4			642 × 1 × 36	576 B
15	ReLU 4			642 × 1 × 36
16	Mean pool 4	642 × 1 × 36	1 × 6 or 1 × 7/1	162 × 1 × 36
17	Convolution 5	162 × 25 × 36	1 × 25/1	162 × 1 × 36	127 KB
18	Batch normalization 5			162 × 1 × 36	576 B
19	ReLU 5			162 × 1 × 36
20	Mean pool 5	162 × 1 × 36	1 × 6 or 1 × 7/1	42 × 1 × 36
21	Fully connected layer 6	42 × 1 × 36	42 × 1/1	1 × 1 × 50	295 KB
22	Batch normalization 6			1 × 1 × 50	800 B
23	ReLU 6			1 × 1 × 50
24	Fully connected layer 7	1 × 1 × 50	1 × 1/1	1 × 1 × 2	408 B
25	Softmax classifier	1 × 1 × 2
Total					1.63 MB

*The size of the mean pool at each node differs according to the number of neighborhood nodes at the node, either 6 or 7.