Bias units in CNN

I want to clarify how a bias unit is added to a 4x4 image?

is b1 added to all 16 elements to the 4x4 matrix or is b1 a 4x4 matrix itself (I’m pretty sure that in week 1 Andrew mentioned b1 as R, not a 4x4 matrix)?

Hello :slight_smile:

Your first guess is the right one. You add the same value to each filter component of the output.
If you have 10 filters in a convolutional layer, you will add 1 different value to each result of the convolution operation

You can see that by printing a model sumary and understanding one conv layer number of parameters

1 Like