Inertial navigation systems (INS) and global navigation satellite systems (GNSS) are often combined to ensure high accuracy navigation. The last generation of inertial measurement units referred to as micro-electro-mechanical systems (MEMS) might also be used in a lot of applications thanks to their relatively low cost. However, the information given by the MEMS is less accurate than with classical INS. In particular, the performance of an integrated GNSS/MEMS navigation systems decreases drastically during GNSS out-ages. This paper studies a neural network based procedure that allows one to compensate this performance loss.