The NTK linear-dynamics approximation accurately predicts the effects of changing the Fourier mapping on training and test loss in the controlled wide-network 1D example; component losses show roughly log-linear convergence. · CiteArk
The NTK linear-dynamics approximation accurately predicts the effects of changing the Fourier mapping on training and test loss in the controlled wide-network 1D example; component losses show roughly log-linear convergence.
来源：paper:§5 p. 6, Figure 3(b–d); §B.1 p. 15, Figure 12
