A shared filter over character embeddings

Input: `^cat$` with two embedding channels
Channel	^0	c1	a2	t3	$4
0	1	0	2	1	0
1	0	1	1	1	0

One filter, reused at every position
Channel	Offset 0	Offset 1
0	1	-1
1	0.5	1

Filter width: 2 · stride: 1 · dilation: 1 · padding: 0 · bias: 0

Window position 0: ^c

(1 × 1) + (0 × −1) + (0 × 0.5) + (1 × 1) + 0 = 2

ReLUGlobal pooling

One output feature map
^c0	ca1	at2	t$3
2	0	2.5	1.5

Pooled value 2.5

max(2, 0, 2.5, 1.5) = 2.5

Illustrative vectors and fixed weights, not a trained model.