Framing RNN as a kernel method: A neural ODE approach