Growing-Neural-Cellular-Automata
Growing-Neural-Cellular-Automata copied to clipboard
Regarding the update rule
I have been following this GitHub repository. Can you explain why you decided to use fully connected layers? Why not use 2D convolutional layers instead?
Thanks. It's actually not me who decided to use FC layers, but the authors (https://distill.pub/2020/growing-ca/) chose to use FC. In my opinion, since the outputs of the Sobel filters are one-dimensional, using FC rather than convolutional layers is quite natural. Although this is not the official implementation, I think I should probably follow the original ideas as best as I can.