Hacker News Reader

It's very pretty of course, but I am curious, the text says,

> that was trained to build textures using neighbour communication only.

How was it trained? on what data? I imagine trained on highly local patterns found in natural images, but I'm trying to imagine what the cost function might have been.

View on HN

https://distill.pub/2020/growing-ca/ - you can either train them to generate expected outcomes

https://distill.pub/selforg/2021/textures/ - this is the paper the authors refer to. Here, the network is trained to maximize activation of another filter of a CNN pretrained network (VGG, for example), which typically correspond to certain textures and shapes. The losses are typical style/gram losses.

Thanks! Makes a lot of sense actually.

Reply on news.ycombinator.com