This observation seems to come up in discussions regarding the hot topic of the https://en.wikipedia.org/wiki/Neural_tangent_kernel https://en.wikipedia.org/wi

This observation seems to come up in discussions regarding the hot topic of the https://en.wikipedia.org/wiki/Neural_tangent_kernel https://en.wikipedia.org/wiki/Neural_tangent_kernel Please correct me if I'm wrong, but I think it means to say that you can in theory conjure specialized kernel methods out of 'infinitely' over-parametrized neural networks. At the moment this all gives unimpressive performance, but it is theoretically promising and could give statisticians interpretable NN-derived models.