DMRS in Neural Receiver

[nnk14102002]

Hello,
I found that the output of the neural receiver has discarded the dmrs symbols because they can't calculate the loss vs label. so does dmrs here really not make much sense when using a neural receiver?
In contrast to the model used in the neural_rx repo, you use a pe vector indicating the distance to the nearest pilot. It serves the purpose of readoutH-est and LLR. Can this be considered one of the improvements to extract information of DMRS?

[SebastianCa]

Hi @nnk14102002,

I’m not entirely sure I understand your question. The neural receiver from the Sionna tutorial utilizes fixed pilots/DMRS for a single user, which remain unchanged after training. In contrast, the multi-user MIMO neural receiver supports dynamic DMRS configurations, as required by 5G, and can be adjusted on-the-fly.

Hope this helps!