I am a new user of matlab deep learning toolbox, however, I would like to build a deeply supervised networks by matlab R2019b. In these networks multiple branches from the successive hidden layers of the main network are used to optian losses (∅d in the equation bellow) at those hidden layers, the training process is required to optimize the final (first term) and intremdiate losses (second term of the equation) in order to fastly converge as shown in figure below! The Objective function is:
∅ = ∅(X,W) + ∑ L* ∅d (X;Wd;w ̃d) + γ(‖W‖^2+∑‖w ̃d ‖^2 )
where ∅ is the entropy loss.
The question is how to merge all these losses in the learning process (training loop)? is there any tutorial could be helpful to address such cases? as the direct way (using trainNetwork) accept only one final outputlayer!