Suppose the input to a Conv2D() layer consists of a 32x32x3 (color) image.

a) How many parameters will there be for a single 3x3 filter (also known as a kernel, also known as a neuron)?  [Don't forget about the bias term]

Since there is only k=1 filter, the height of the filter is h=3, the width of the filter is w=3, there are d=3 input channels from the previous layer, and a filter will only have b=1 bias parameter, a single filter will have k * (h * w * d + b) = 1 * (3 * 3 * 3 + 1) = 28 parameters.

b) How many parameters in total for the Conv2D() layer if there are 32 filters?

A set of k=32 filters will have k * (h * w * d + b) = 32 * (3 * 3 * 3 + 1) = 896 parameters.

The following code can be used to check your answer:

from tensorflow.keras import models, layers
model = models.Sequential([ layers.Conv2D(filters = 32, kernel_size = (3, 3), input_shape = (32, 32, 3)) ])
model.summary()