If you are a person who has made many attempts in neural networks, especially if you design many structures yourself instead of copying them, then when constructing networks in practical applications, I believe that there is often such a requirement: some parameters need not to be updated.
Such scenarios may appear in trained word vectors, pre-trained network modules with parameters, and the bottom layer where reliability assignment is prone to problems.
How to achieve parameter differentiation in Lasagne?
There are two methods here:
1. Filter by assigning different labels to parameters when initializing the network parameters.
2. Modify the network parameter list directly to avoid the participation of some parameters in the update process.