Curious on possible application of Neural ODE?

[jlevy44]

I'm just curious. This is coming from someone with a naive understanding of your methodology. I was reading through the STGCN paper, and came across your method. Seems like parameter reduction and filter localization is done through restriction of the kernel to a polynomial.

Is it possible, in anyway, to replace this filter with, and again I am a bit naive here, with a neural ordinary differential equation? That is, representing the filter as a diffeq rather than a polynomial and learn those set of diffeq parameters? Would such parameters help reduce the complexity of the model?

This is a point of curiosity, just looking for enlightenment. If this is possibly, than it seems the neural ODE framework could be extended to GCN.