Question

A 10x10 image is the input to a convolutional neural network (CNN). The first hidden layer of the CNN is comprised of 4 3x3 filters. Assuming zero-padding and bias, what is the dimensionality of the output from the first hidden layer and the number of parameters that are estimated from this layer?

A 10x10 image is the input to a convolutional neural network (CNN). The first hidden layer of the CNN is comprised of 4 3x3 filters. Assuming bias but not zero-padding, what is the dimensionality of the output from the first hidden layer and the number of parameters that are estimated from this layer?

Answer 1

Formula =

Output = ([Input + 2 * padding - kernal_size]/stride)+1

Parameter = ((width of the filter * height of the filter * number of channels_in_prev_layers+1)*number of filters)

NOTE-: 1 is added for the bias

Ans 1)

([10+2*0-3]/1)+1     [By default stride is 1]

7+1 = 8

Output Shape = (8, 8, 4) where 4 is the number of filters

Parameters = (3*3*1+1)*4

                   = 40

Ans 2)

([10+2*p-3]/1)+1     [By default stride is 1]

(7+2p)/1+1 = 8+2p

Output Shape = (8+2p, 8+2p, 4) where 4 is the number of filters

Parameters = (3*3*1+1)*4

                   = 40