-2

The result of the convolution operation is multiple subsets of data are generated per kernel. For example if 5 kernels are applies to an image of dimension WxDx1 (1 channel) then 5 convolutions are applied to the data which generates a 5 dimensional image representation. WxDx1 becomes W'xD'x5 where W' and D' are smaller in dimension that W * D

Is the fact that each kernel is initialised to different values prevent each kernel from learning the same parameters ? If not what prevents each kernel learning the same parameters ?

If the image is RGB instead of grayscale so dimension WxDx3 instead of WxDx1 does this impact how the kernels learns patterns ?

blue-sky
  • 51,962
  • 152
  • 427
  • 752

1 Answers1

1

As you already mentioned, the sole fact of differing what Kernels learn is due to the random initialization of the weights in the beginning.

A great explanation is delivered here and also applies for the convolutional kernels in CNNs.
I regard this as distinct enough to not highlight it as a duplicate, but essentially it works the same.

dennlinger
  • 9,890
  • 1
  • 42
  • 63