And here is the image of my poster:

With this method, I can generate CNNs that have 2 million parameters, with architectures that haven't been exposed to the local rule during its training.

I'm also extending the general neighborhood definition from my previous work on this to include CNN architectures, and I'm presenting the results of architecture generalization of the local rule on both MLPs and CNNs.

In MetaNCA I'm using an NCA to update the weights of a neural network to be good at a particular task. The updates are based only on the local neighborhood of weights. In this poster, I'm presenting a Weight Transformer to aggregate the messages of these neighbors.

Labels for the communities could be helpful. What defines a link between two words in these diagrams?