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

https://distill.pub/2020/growing-ca/ https://distill.pub/2020/growing-ca/ - you can either train them to generate expected outcomes https://distill.pub/selforg/2021/textures/ 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.