Convolution Kernels

A kernel is a small set of weights reused across positions.

In one dimension, a kernel can slide over a sequence:

input:  [1, 2, 3, 4]
kernel: [1, 0, -1]

At the first position, it reads [1, 2, 3]:

1*1 + 2*0 + 3*(-1) = -2

At the next position, the same kernel reads [2, 3, 4]:

2*1 + 3*0 + 4*(-1) = -2

The same weights are reused. That reuse is what makes convolution different from giving every position its own separate weights.