In the modern world, the use of Internet based communications has become a daily activity for a big part of the population. In particular, mobile communications are experiencing a relentless growth in the number of users, which has become an issue as they must be allocated in the channels with the limited available time and frequency resources in the physical layer. This is why Non-Orthogonal Multiple Access (NOMA) schemes are being used in 5G mobile communications and are expected to be a key factor in the future 6G communications. Another factor to take into account is the evolution in technologies such as Artificial Intelligence (AI) and Machine Learning (ML) that, as proven in literature, can perform tasks in the telecommunications world. NOMA will play an important role in 6G communications, while AI/ML has proven to play a key role in many layers including the physical layer, to improve the performance. In particular, constellation design to optimize BER vs SNR performance has proven to be feasible via Deep Learning (DL). In this study, we focus on neural design of demappers for an uplink NOMA system, to demonstrate the performance advantages of these systems against classical demapping techniques. The traditional demapping techniques to evaluate are the Successive Interference Cancellation (SIC) and the Joint Demapper. On the other hand, the SICNet demapper , a ML version of the traditional SIC, has been adapted to the uplink NOMA system. The comparison of the performance of all the demappers has been done with BER-EbN0 curves. The results show that the Joint demapper is the most optimal technique in a two-user uplink system. Meanwhile, the SICNet demapper has proven to outperform the SIC demapper in most of the studied scenarios.