Thread by @quasimondo

I've created an experimental GAN architecture I call #RecuResGAN or "Recursive-Residual GAN" and I am pretty astonished that:
- it works at all
- how well it works across a pretty wide range of scales.
- it is just 15% the size of a comparable #pix2pixHD model

Of course I did not google the concept of recursive neural networks before I started this experiment and enjoyed the illusion of being very innovative here for a whole day:
en.wikipedia.org/wiki/Recursive…

The principle is pretty simple: in a classic residual architecture you chain several residual blocks behind each other (in #pix2pixHD the default is 9 blocks), what I do in #RecuResGAN is to use a single block, but loop 9 times over it, feeding its output back into its input.

The same goes for the down- and up-convolution modules, only here you cannot reduce or increase the amount of filters within a block so you have to compromise a bit how many you use in order not to run out of memory with the accumulated gradients.

So my theory why it seems to be relatively scale invariant is particular because of those recursive up- and down convolutions where a block has to handle all the features at various scales and thus becomes kind of fractal.