Current polarimetric calibration techniques derive the instrument modulation matrix by numerically fitting an instrumental model to measurements of a set of calibration Stokes vectors. These techniques are typically limited to an error on the retrieved normalized Stokes Q, U and V parameters, in the $1\times10^{-2}$ to $1\times10^{-3}$ range. This error commonly increases when the instrument response varies considerably across its field of view and/or when instrumental effects are present, which are not included in the assumed calibration model, such as camera non-linearity. We propose a new technique to calibrate Stokes polarimeters based on a model composed of fully-connected, multi- layer neural networks (NN's). This model is trained to learn the instrument modulation matrix given a set of input parameters, using the same calibration data that is acquired for the current techniques. The main advantage of our NN-based approach is its flexibility to incorporate instrumental effects for which no accurate model is available and, possibly through the fusion of data from other types of relevant calibrations, obtain a more accurate instrument response. We present a preliminary result of our model performance using synthetic data, where the ground truth is known.