Future progress with GLN

[aisurfer]

Hello!
Guys do you figure out why there is no progress with GLNs research?
Does it fail on most tasks? Or just hard to beat a SOTA?
Why authors or other Deepmind researchers do not continue their work on it?

[AlexKuhnle]

Hi, sorry for the very late response, didn't get a notification for this. Not aware of any further developments, and if I would have to guess it's that it just doesn't work that well in practice. My impression when implementing GLN was that in particular the number of parameters wouldn't scale well when moving to bigger architectures and less "trivial" tasks than MNIST.