Learning normal form autoencoders for data-driven discovery of universal, parameter-dependent governing equations.