Phase-modulated dispersion interferometry is an increasingly popular method of measuring the line-integrated electron density in magnetic confinement fusion devices due to its inherent vibration compensation in combination with the high resilience to rapid density changes. At present, two quadrature methods of extracting the phase from a temporal interferogram of a phase modulated DI are known, both of which have been shown to exhibit strong phase nonlinearities, which could be prohibitively large in fusion reactor size devices (Brunner et al 2022 Rev. Sci. Instrum.93 023506). These errors are not fixed and hence, i.e. they change over the course of the day, which makes static compensation schemes difficult. In this work we present a neural network based approach to extract the phase information from a temporal dispersion interferogram using a mathematically regulated auto-encoder. No assumptions are being made on the underlying physics/optics of the data avoiding training the network with false assumptions. The training approach also avoids the common concern of extrapolation entirely, i.e. how does the network behave when it sees unseen data. We also show that the network can be trained without using plasma data. The network is shown to extract quadrature components from the interferogram with significantly reduced systematic quadrature distortions as well as lower noise and higher bandwidth than the currently established methods. It is also shown to effectively reduce the nonlinearity drifts occurring, when trained on larger datasets. The method presented here is not limited to dispersion interferometers, but should be applicable to any quadrature component based phase measurement, e.g. radiometry, reflectometry, radar or digital radio.