Why use h_hat interpolated by “nn” as input to the network

[modestlyh]

I noticed that instead of using only LS estimates, intitial LS channel estimation still uses nn interpolation, which by definition should use only LS estimates, because if interpolation is done, that is equivalent to the network not including the channel estimation piece, I don't know if I'm understanding correctly，but the dimension seems to be an issue again if only LS estimates are used

[SebastianCa]

Hi @modestlyh,
That's right, there is an initial LS estimation stage. This is useful in multi-user settings and avoids the need to feed (potentially dynamic) pilots during inference (as the value of pilots in 5G NR is dependent on the slot id).

Note that nn stands for nearest neighbor interpolation, meaning the LS estimates at the pilot positions are simply copied to fill the entire resource grid. Compared to the overall complexity of the neural receiver this adds negligible complexity.

[modestlyh]

@SebastianCa Thanks a lot！

[modestlyh]

@SebastianCa I would like to ask, has it been tried to train at a high signal to noise range, say 10-25db, and then evaluate at a lower signal to noise range with good results?

[SebastianCa]

The performance is typically better if the training is done over the entire SNR range, i.e., including low SNR values as well.