In order to estimate reliable models from noisy input-output data, system identification techniques usually require that the data be generated by a process with a fading memory. Non-equilibrium systems such as neuronal and chaotic models lack a fading memory. Their identification is challenging, in particular in the presence of input noise. In this thesis, we propose a methodology based on the prediction-error method for the identification of neuronal systems subject to input-additive noise. We build on the fundamental observation that while a neuronal model does not have a fading memory, it can be transformed into a fading memory system by output feedback. Our ideas can be generalized to any non-equilibrium system sharing this property. At the core of the methodology is the use of output feedback in experiment design. We provide a theoretical justification for this design choice, which has been exploited in neurophysiology since the invention of the voltage-clamp experiment. To investigate the problem of feedback for identification, we first address the estimation of simple non-equilibrium systems in Lure form, and show that feedback allows estimating the nonlinearity in a static experiment. We then address the estimation of conductance-based models. Assuming that an informed choice can be made on the elements of the model structure, we show that consistent parameter estimates can be obtained when noise is only present at the system input. Finally, we approach the problem from a black-box perspective, and propose identifying the neuronal internal dynamics using a universal approximator with Generalized Orthogonal Basis Functions.