←back to thread

354 points misonic | 1 comments | | HN request time: 0.201s | source
Show context
samsartor ◴[] No.42468798[source]
GNNs have been a bit of a disappointment to me. I've tried to apply them a couple times to my research but it has never worked out.

For a long time GNNs were pitched as a generalization of CNNs. But CNNs are more powerful because the "adjacency weights" (so to speak) are more meaningful: they learn relative positional relationships. GNNs usually resort to pooling, like described here. And you can output an image with a CNN. Good luck getting a GNN to output a graph. Topology still has to be decided up front, sometimes even during training. And the nail in the coffin is performance. It is incredible how slow GNNs are compared to CNNs.

These days I feel like attention has kinda eclipsed GNNs for a lot of those reasons. You can make GNNs that use attention instead of pooling, but there isn't much point. The graph is usually only traversed in order to create the mask matrix (ie attend between nth neighbors) and otherwise you are using a regular old transformer. Often you don't even need the graph adjacencies because some kind of distance metric is already available.

I'm sure GNNs are extremely useful to someone somewhere but my experience has been a hammer looking for a nail.

replies(5): >>42468874 #>>42468882 #>>42469313 #>>42469395 #>>42472618 #
1. igorkraw ◴[] No.42469313[source]
GNNs are useful at least in one case, when your data a set of atoms that define your datum through their interactions, specifically a set that is that is high cardinality (so you can't YOLO it with attention) with some notion of neighbourhood (i.e. geometry) within your set (defined by the interactions) which if maintained makes the data permutation equivariant, BUT you can't find a meaningful way to represent that geometry implicitly (for example because it changes between samples) => you YOLO it by just passing the neighourhood/interaction structure in as an input.

In almost every other case, you can exploit additional structure to be more efficient (can you define an order? sequence model. is it euclidean/riemanian? CNN or manifold aware models. no need to have global state? pointcloud networks. you have an explicit hierarchy? Unet version of your underlying modality. etc)

The reason why I find GNNs cool is that 1) they encode the very notion of _relations_ and 2) they have a very nice relationship to completely general discretized differential equations, which as a complex systems/dynamical systems guy is cool (but if you can specialize, there's again easier ways)