Would it be possible to have an LSTM that is followed by two output layers, where each output layer computes a different representation and is followed by two different loss functions (i.e. where the result of the first output layer will be fed into a cross entropy loss function and the result of the second output layer is fed into L2 loss function and the losses are back-propagated together)

Would there be any obvious problems with this?

Also can anyone point me to a paper that is using something like this?

  • $\begingroup$ In general, I think you can do this; however, I'm not super familiar with recurrent architectures that do what you're describing, but I can think of architectures in computer vision that do something similar to what you're saying with a very clear motivation as to why they use multiple losses. I'm unsure how this would function in the recurrent setting; could you give more info on what your task is and why you're considering this? $\endgroup$ – tchainzzz Jun 9 at 1:27

Your Answer

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

Browse other questions tagged or ask your own question.