Using MSE for Reverse Loss in Gaußian Mixture Model implementation

[StefanHauser]

Hello,

I'm currently doing a project work on INNs during my master study at university of applied science in Karlsruhe.

As an orientation I was using the toy implementation of the Gaußian Mixture Model.
I got strange results during my first approach as the reverse loss was rising extremely every few epochs - both on train and test data.

Further into the code, I've seen that the reverse_orig_loss is implemented with the MSE loss:

    l_rev += lambd_predict * loss_fit(output_rev, x)

You write in the paper that the reverse loss has to be implemented with MMD. When I've changed the reverse loss to MMD, I got way better results and the loss is shrinking really well.

Is there a reason why you've used MSE at this point?

Thanks in advance!