Thanks Taro. I am now aware of over-squashing problems arising in classical CNNs (the growth of the number of neighbours is also polynomial vs exponential in "small-world" graphs). However, these phenomena were first observed in seq2seq models, so they are not specific to graph neural networks.