The final kernel is an integral taken over the path of the parameters during the gradient descent. Basically what they're doing is just y = y0 + \int dy/dt dt

The final kernel is an integral taken over the path of the parameters during the gradient descent. Basically what they're doing is just y = y0 + \int dy/dt dt and then rephrasing this equation as a kernel machine. Which I guess should also work for other forms of learning where you slowly change the parameters.