0
$\begingroup$

So I am trying to implement a specific CNN called a U-net. It states in page 3 that it doesn't have a fully connected layer.

Till then I understood CNN to have two stages; 1. Convolution, where the kernels are learnt and features are extracted and 2. where the extracted features are flattened and a Fully Connected Layer (FCL) is applied where weights and biases are learnt. This U-Net though has gotten me slightly confused and I wanted to make sure whether CNN need to be fully connected or not.

If they do not, then where does the 'neuron behaviour' come from in CNN? In what sense are they Neural Nets then?

Improve this question
$\endgroup$
0
$\begingroup$

I think this depends on the use-case. CNNs are often applied in an image classification setting, where as you mention, the convolution layers are often followed by FCLs. With U-net, on the other hand, the result is another image showing a segmentation map, so no FCLs are used.

The 'neuron behaviour' in U-net comes from ReLU activation functions, which are applied after the convolution filters. (See the legend of the architecture picture here: https://lmb.informatik.uni-freiburg.de/people/ronneber/u-net/)

Improve this answer
$\endgroup$
2
  • $\begingroup$ Do you mean to say that the 'neural' aspect is only because a feature map is active if certain conditions are met (in ReLU, being more than 0). NOT due to the interconnectedness of nodes? $\endgroup$
    – Udb
    Mar 27 '20 at 15:40
  • $\begingroup$ The 'neural' aspect comes from the non-linearities. Only increasing interconnectedness without adding non-linearity severely limits what any NN can do. It can even be shown that a basic NN of any depth without any activation functions can be reduced to a single layer NN (since without them, only linear matrix multiplications remain) and is basically just a linear regression. In CNNs the convolution operations themselves are also linear mappings, so you need the activation functions to introduce non-linearity. $\endgroup$
    – phnx
    Mar 27 '20 at 15:56

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.