If we consider a CNN that takes in $32 \times 32$ grayscale images and has a single convolution layer with 3 different $3 \times 3$ convolution filters (without or with boundary padding). I wonder whether we can treat this CNN as a feed-forward NN with each pixel as input and posting constraints on the weights in the hidden units? If so, what could be the constraints?
Asked
Active
Viewed 31 times
