A method of defining and setting a nonlinear signal processing of a hearing device, e.g. a hearing aid, by machine learning is provided. The hearing device being configured to be worn by a user at or in an ear or to be fully or partially implanted in the head at an ear of the user, the method comprising providing at least one electric input signal representing at least one input sound signal from an environment of a hearing device user, determining a normal-hearing representation of said at least one electric input signal based on a normal-hearing auditory model, determining a hearing-impaired representation of said at least one electric input signal based on a hearing-impaired auditory model, determining optimised training parameters by machine learning, where determining optimised training parameters comprises iteratively adjusting the training parameters, and comparing the normal-hearing representation with the hearing-impaired representation to determine a degree of matching between the normal-hearing representation and the hearing-impaired representation, until the degree of matching fulfils predetermined requirements, and, when the degree of matching fulfils the predetermined requirements, determining corresponding signal processing parameters of the hearing device based on the optimised training parameters. A hearing device is further provided.