CliqueNet and backward connection neural network


I would like to implement the CliquetNet neural networks with fully connected dense with Flux:


The network has some backward connection, I do not know the best way to implement the connection from the Flux.SkipConnection and the Flux.Parallel, as I understand only the forward connection is possible.

Does anyone have an idea?


As a first step, you could try porting the official implementation at GitHub - iboing/CliqueNet: Convolutional Neural Networks with Alternately Updated Clique (to appear in CVPR 2018). If Flux’s built-in layers aren’t appropriate, don’t be afraid to create custom ones like the authors did.