salt-nlp / adaptive-compositional-modules Goto Github PK

In the paper, it is said that the initialization of c_{k+1} is -c, and the others are c (c> 0) can increase the tendency to reuse old modules. I have some questions about this:
(1) Is c a constant in the training process? Can c_k be understood as the weight corresponding to the k-th module?
(2) I understand that this can inhibit the effect of new adapter, but how does the model decide whether to add a new adapter in the end? How do you decide which adapter to share for similar tasks?
(3) Is the adapter for the old task constantly updated or fixed?
(4) Whether task identifiers are required for inference？

Looking forward to your reply， best wishes