Exchangeability and Kernel Invariance in Trained MLPs