Recently, I have read some cnn papers, including fcn method.
In the traditional cnn methods, every layers eg. conv_layers, fc_layers, are needed to train to optimize the weights and filter kernels.
I'm wonder a thing(here is the thing I can't figure out and my question is), should the filters used in deconvolution layers need to be trained? Or they just use the same filters of the previous corresponding convolution layers and no need to train?
Thanks in advance!
The weights of the filters in the deconvolution layer do need to be trained. Just as in another other layer would, they have a forward pass and a backward pass.