Sander Dieleman

I'm actually mistaken -- most of the logic should be in `Layer.add_param`, so we could modify that function to store the initializers somewhere. So it is a more local change...

All good point, clearly this issue can use some more brainstorming so we get it right :)

as I understood it, they just add a bunch of all-zero feature maps. This basically means that all the additional features in excess of the number of input features are...

Do they have a stride 2 1x1 convolution in the paper? Don't they do that only for the 3x3 convolution? I don't know, I haven't checked. But a stride 2...

Cool! This would be a really great addition to our Recipes :)

Although pickling has its issues, are we sure we want to block it completely? There might be some use cases? I guess a warning would be most appropriate.

Unfortunately the current interface for recurrent layers is not particularly modular. We've been talking for a long time about fixing this, but so far the plans haven't really materialised. Have...

It's difficult for me to judge whether the use cases are important enough to warrant the additional complexity here. Overall I think we should follow "explicit is better than implicit",...

I'm not sure how feasible it is to implement this as a function in lasagne.updates, to be honest. I'm not super familiar with the method, but doesn't it make extensive...

I think a better strategy would be to create a sister library for TensorFlow, that provides largely the same API as Lasagne (where possible), but returns TensorFlow objects where Lasagne...