Same goes for ImageNet: These GNNs are getting only 50% top-5 accuracy, well below SOTA. But I find that figure impressive: It implies that I can concoct a new

Same goes for ImageNet: These GNNs are getting only 50% top-5 accuracy, well below SOTA. But I find that figure impressive: It implies that I can concoct a newfangled architecture for visual recognition, put it through one of these GNNs, and begin training at ~50% top-5 accuracy. Also, AFAIK this is the first effort of its kind; I would expect the figures to improve over time, as usual.