Hi Everyone - I’m starting to think about using differential learning with a UNet. I’ve only seen this applied to traditional classification problems and I was wondering if the architecture of the UNet - and in particular the skip connections should change how I approach the setup.
Are there any subtleties (skip-connection related or otherwise) I should be thinking about or should i just proceed in the standard fashion - freezing early layers and then slowly increasing the learning rates as I progress through the later layers?