//=time() ?>
hi-res texture synthesis, i.e. contentless style transfer. here is one based on Kandinsky's Composition 8 (1600x1260 version at https://t.co/QjaHr3eVa7)
Berlin friends, new workshop: learn how to reverse neural nets to make your own deep dreams. Nov 25, @spektrumberlin
https://t.co/PMMpwZZ8x5
thanks to @JakeElwes for making this supercut of me saying nothing but random numbers 🤣
i.e. a "DaDaDa Ta" https://t.co/IxJik5VIG1
some more neural synthesis/deepdream experiments... masks for new channels radiating outward from the center
examples of laplacian normalization + mixing gradients. sometimes channels blend features, other times negotiate over space and interweave
@mtyka here's three channels arranged in overlapping rings. the aesthetics between channels can be unpredictable. lot of territory to explore
CycleGAN "failure cases" are as interesting as the successes. ML papers need more blooper sections https://t.co/qHZi0sfH6b