I'm having trouble understanding how bias is added to the feature extraction convolution. I've seen people either refer to the bias as a single number that changes per filter or the whole matrix that is the size of the output. Here is what I mean:

enter image description here

enter image description here

  • $I$ is the input single-channel image.
  • $F$ is the filter.
  • $b$ is the bias.
  • "Izhod" means "output".

Which is actually the correct bias used in CNN?

  • $\begingroup$ Hello. Where did you take this diagram from? $\endgroup$
    – nbro
    Jun 12, 2021 at 2:11
  • $\begingroup$ I drew it myself. The idea that bias is a single number and not a matrix came from cs231n.github.io/convolutional-networks under the Convolution demo where Bias is displayed as a 1x1x1 matrix. However the idea that the bias might actually be a whole matrix that is the size of the output came from this video youtu.be/Lakz2MoHy6o?t=343 at the 5:00 mark. $\endgroup$
    – kranj doo
    Jun 12, 2021 at 11:42


You must log in to answer this question.