これこれ

Every Model Learned by Gradient Descent Is Approximately a Kernel Machine
https://arxiv.org/abs/2012.00152