PaperlayerView exact revision on arXiv

Source-linked AI summary

Cross-Site Transfer and Spatial Forgetting in OTA-Trained Neural Receivers

Riku Luostari, Dani Korpi, Harri Holma, Olav Tirkkonen

arXiv:2610.02577v2eess.SP

TL;DR

The paper asks how well OTA data transfers across sites and what deployment-site adaptation costs elsewhere. It compares cross-site training protocols and finds that the value of deployment data depends on operating point, with finetuning trading other-site retention for a larger local gain.

  • Problem

    The study asks how the value of OTA data depends on where it is collected, including whether cross-site transfer can make deployment-site measurements worthwhile.

    Source: §I · p. 1

  • Method

    The study measures transfer from seven source sites to a distinct deployment site and compares joint training with finetuning on identical measurements.

    Source: §I · p. 1

  • Results

    At the higher-rate point, finetuning yields larger deployment-site sensitivity but about 0.1 dB of spatial forgetting, while remaining about 0.4 dB ahead of simulation-only training on the aggregate other-site test set.

    Source: §VI · p. 6

  • Takeaways & Limitations

    Use other-site measurements as a transferable starting point, add deployment-site data especially when higher-rate sensitivity matters, and choose joint training or finetuning according to the priority placed on other-site retention versus local sensitivity.

    Source: §VI · p. 6

    The relative value of data origin and adaptation protocol depends on the operating point, so the trade-off is not uniform across the evaluated points.

    Source: §V · p. 5

Abstract

from arXiv · show

Most published neural-receiver studies train on synthetic channel models, although real over-the-air measurements can improve performance. We examine how much of this benefit transfers from other sites to a distinctly different deployment site and what adaptation to that site costs elsewhere. The study uses a dataset from a 5.88 GHz measurement campaign covering eight sites and four environment types. An industrial site measured entirely from a vehicle is withheld as a deliberate transfer stress test. One receiver architecture is trained under four regimes: simulation alone; seven other sites; all sites jointly; and other-site training followed by deployment-site finetuning. The last two use identical measurements through different optimization protocols. Across five independently trained seeds, sensitivity gains are evaluated at uncoded bit error rate (BER) targets of $10^{-1}$ and $10^{-2}$, representative of moderate- and higher-rate coded links, respectively. On the deployment site, training with other-site data rather than simulation improves sensitivity by 0.3 dB at $10^{-1}$ and 0.1 dB at $10^{-2}$. Adding deployment-site data contributes less than 0.1 dB at $10^{-1}$; at $10^{-2}$, total gains over simulation reach 0.4 dB with joint training and 0.6 dB with finetuning. Joint training shows no measurable aggregate loss on the other-site test set, whereas finetuning yields the larger deployment-site gain but reduces other-site sensitivity by about 0.1 dB at $10^{-2}$, indicating spatial forgetting.

I. INTRODUCTION

The study asks how far OTA training transfers across sites and whether deployment-specific adaptation costs sensitivity elsewhere. It tests seven-site pretraining against a deliberately distinct held-out site and compares joint training with finetuning on matched data.

I. INTRODUCTION

  • The campaign spans eight sites and four environment types, and holds out an industrial site of scattered steel halls measured entirely from a vehicle as a transfer stress test.The four training regimes are simulation alone, other-site data, joint other- and deployment-site data, and other-site training followed by deployment-site finetuning.Source: §I · p. 1
  • The study quantifies transfer from seven source sites to a deliberately different deployment site, with other-site data contributing more at the moderate-rate point and deployment-site data more at the higher-rate point.The study also compares joint training against finetuning using identical measurements and evaluates run-to-run variation across five paired seeds.Source: §I · pp. 1–2
  • Joint training shows no measurable aggregate loss on the other-site test set, whereas finetuning produces measurable spatial forgetting while still outperforming simulation-only training there by about 0.4 dB.The comparison uses identical deployment-site measurements across the two adaptation protocols.Source: §I · p. 2

II. SYSTEM MODEL

The system model uses a SISO OFDM link with 16QAM and known pilot positions. Results are reported against a classical LS receiver whose stages are data-independent.

II. SYSTEM MODEL

  • The received signal at subcarrier j and symbol i is modeled from the transmitted symbol, complex channel gain, and additive white Gaussian noise.The system uses one transmit and one receive antenna, 14 OFDM symbols per slot, and 16QAM.Source: §II · p. 2
  • The OFDM setup retains the parameters from [5], with known pilots at specified resource-element indices and the remaining active subcarriers carrying data or pilots.The pilot set is denoted P; guard bands and the DC subcarrier are excluded.Source: §II · p. 2, Table I · p. 2
  • The LS RX estimates the channel at pilot positions, interpolates the remaining resource elements, applies zero-forcing equalization, and demaps symbols to per-bit LLRs.The baseline is chosen for mathematical tractability and is identical across training arms and sites; LMMSE would require an assumed channel covariance.Source: §II · p. 2

III. METHODS

The experiments use Hybrid CM DeepRx, a receiver that maps the received resource grid and known pilots directly to per-bit LLRs. Its hybrid block switches into the delay domain to expose localized multipath structure.

III. METHODS

  • Hybrid CM DeepRx maps the received grid and known pilots directly to per-bit LLRs without forming an explicit channel estimate.The receiver uses six ConvMixer blocks and has 531,972 trainable parameters.Source: A. Hybrid CM DeepRx · p. 2
  • The first block switches to the delay domain for depthwise and pointwise mixing before returning to the resource grid, while the remaining blocks operate on that grid.The design widens the frequency kernel relative to CM DeepRx and uses complex multiplication in the pointwise stage, which can represent phase rotations.Source: A. Hybrid CM DeepRx · p. 2
  • Fig. 2 summarizes the receive chain and Hybrid CM DeepRx architecture, including synchronization, cyclic prefix, convolution, FFT, and IFFT stages.The caption expands Sync., CP, and conv as synchronization, cyclic prefix, and convolution.Source: Figure 2 · p. 2

B. Measurement Campaign

The measurement campaign uses a real 5.88 GHz SDR link across eight sites, with varied environments and mobility. Its industrial site is the sole industrial and entirely vehicular site, selected in advance as the deployment-site holdout.

B. Measurement Campaign

  • The campaign uses modified PlutoSDRs at 5.88 GHz, with the receiver mounted at base-station sites and the transmitter handheld or vehicle-mounted.The received signals include practical impairments often omitted from synthetic channel models, including amplifier nonlinearity, I/Q imbalance, phase noise, and timing and frequency errors.Source: B. Measurement Campaign · pp. 2–3
  • The eight sites cover urban, suburban, rural, and industrial environments, while the industrial site consists of scattered steel halls and was measured only from a vehicle.That site is the only industrial site and the only entirely vehicular site; it was selected before model training or evaluation.Source: B. Measurement Campaign · p. 3
  • CRC-validated payload labels support training and BER evaluation, with fewer than 1% of slots discarded after CRC failure.The study fixes the modulation and coding scheme and evaluates uncoded-BER sensitivity gains; testing other schemes would require additional OTA data.Source: B. Measurement Campaign · p. 3

C. Training Arms and Dataset Splits

The study compares four training arms to separate cross-site transfer, the value of deployment-site data, and optimization protocol. Data are split by site, with slot rows randomly assigned within each site.

C. Training Arms and Dataset Splits

  • Slot rows are randomly assigned within each site, and only 1,000 deployment-site measurements enter training; neither test set informs training decisions.Table II lists the resulting measurement counts.Source: C. Training Arms and Dataset Splits · p. 3, Table II · p. 3
  • The four arms use simulation-only training, seven other sites, all sites jointly, or other-site training followed by deployment-site finetuning.The other-site versus simulation comparison measures transfer; all-site versus other-site measures added deployment-site data; all-site versus finetuned compares protocols on identical measurements.Source: C. Training Arms and Dataset Splits · p. 3
  • Measured-data arms add AWGN that lowers SINR by a uniformly sampled amount up to 1 dB; simulated samples are not augmented.Each simulated sample uses a freshly drawn channel, noise realization, and SINR.Source: C. Training Arms and Dataset Splits · p. 3
  • The all-site and finetuned arms use the same 28,946 measurements but differ in optimization: joint training mixes deployment data throughout, while finetuning uses it in a second stage.The finetuned model starts from the other-site checkpoint and does not see deployment-site data twice.Source: C. Training Arms and Dataset Splits · p. 3

D. Site-Specific Finetuning

Finetuning adapts the other-site model using deployment-site measurements, and sensitivity gain is assessed at two uncoded BER targets representing different coded-link rates.

D. Site-Specific Finetuning

  • Finetuning uses deployment-site data exclusively for 10,000 iterations at batch size 16 and a fixed learning rate of 4×10^-6, with evaluation at the final checkpoint.The adaptation set contains 1,000 measurements; its size and learning rate were selected without either test set.Source: D. Site-Specific Finetuning · p. 3

E. Evaluation Metrics

  • Sensitivity gain is the SINR decrease needed to reach a target uncoded BER, estimated by logarithmic interpolation of binned BER curves.The BER targets 10^-1 and 10^-2 represent moderate- and higher-rate coded links, respectively.Source: E. Evaluation Metrics · pp. 3–4

F. Replicates and Statistical Analysis

Five paired seeds quantify transfer and run-to-run variation. Other-site training improves deployment-site sensitivity over simulation at both evaluated BER targets, while the same model shows a gap between test sites.

F. Replicates and Statistical Analysis

  • Each of five independent seeds defines a full pipeline replicate, with finetuning paired to its corresponding other-site model and differences computed within seed.Reported differences are means and standard deviations across the five paired seeds.Source: F. Replicates and Statistical Analysis · p. 4

A. Transfer of Over-the-Air Data Across Sites

  • At the deployment site, the other-site model yielded 0.29 ± 0.03 dB more sensitivity gain than the simulation-only model at 10^-1, and 0.12 ± 0.05 dB more at 10^-2; values are five-seed means ± across-seed standard deviations.Source: A. Transfer of Over-the-Air Data… · p. 4, Table III · p. 4
  • The other-site model’s sensitivity gain is 2.26 ± 0.01 dB on the other-site test set versus 1.91 ± 0.02 dB on the deployment site at 10^-1, and 2.58 ± 0.02 versus 2.44 ± 0.05 dB at 10^-2.The unseen-site gaps are 0.35 dB and 0.14 dB, respectively; nearly identical SINR distributions do not explain them, and local LS RX comparisons reduce but do not eliminate test-set difficulty differences.Source: A. Transfer of Over-the-Air Data… · p. 4, Table IV · p. 4

B. Value of Deployment-Site Data

Deployment-site measurements add little sensitivity at 10^-1 but more at 10^-2. Finetuning gives a larger deployment-site gain than joint training, despite using the same measurements.

B. Value of Deployment-Site Data

  • At 10^-1, joint training improves deployment-site sensitivity by 0.04 ± 0.01 dB and finetuning by 0.07 ± 0.01 dB over the other-site model, for total gains over simulation of 0.33 ± 0.03 and 0.36 ± 0.03 dB.Deployment-site data contributes a larger increment at 10^-2 than at 10^-1.Source: B. Value of Deployment-Site Data · p. 4
  • At 10^-2, joint training improves deployment-site sensitivity by 0.29 ± 0.06 dB and finetuning by 0.47 ± 0.01 dB over the other-site model, for total gains over simulation of 0.41 ± 0.04 and 0.59 ± 0.04 dB.The arm ordering is identical in every seed at both operating points.Source: Figure 5 · p. 4, B. Value of Deployment-Site Data · p. 4
  • With identical measurements, finetuning adds 0.03 dB over joint training at 10^-1 and 0.18 ± 0.05 dB at 10^-2.The across-seed standard deviation for the 10^-1 difference is below 0.01 dB.Source: B. Value of Deployment-Site Data · p. 4

C. Adaptation versus Forgetting

Finetuning on deployment-site data improves deployment-site sensitivity but reduces sensitivity on the other sites, while joint training shows no measurable aggregate loss there. The finetuning curves improve locally and decline elsewhere, then flatten toward the end of the prescribed budget.

C. Adaptation versus Forgetting

  • Joint training changes other-site sensitivity by 0.00 ± 0.01 dB at 10^-1 and −0.01 ± 0.02 dB at 10^-2 relative to other-site training.These paired differences indicate no measurable aggregate loss on the other-site test set.Source: C. Adaptation versus Forgetting · p. 5
  • Finetuning lowers other-site sensitivity gain by 0.04±0.01 dB at 10^-1 and 0.11±0.04 dB at 10^-2, yet remains about 0.4 dB ahead of simulation-only training on the other-site test set.The loss is measured relative to the other-site starting model after finetuning with 1,000 deployment-site measurements and a fixed learning rate of 4 × 10^-6.Source: C. Adaptation versus Forgetting · p. 5
  • At 10^-2, repeated finetuning runs reach 0.50 ± 0.03 dB at the deployment site and −0.09 ± 0.01 dB at the other sites after 10,000 iterations.By 5,000 iterations, the corresponding changes are 0.43 ± 0.02 dB and −0.07 ± 0.01 dB; the final 5,000 iterations add about 0.07 dB locally and remove about 0.02 dB elsewhere.Source: Figure 7 · p. 5, C. Adaptation versus Forgetting · p. 5
  • The prescribed 10,000-iteration budget was not selected as an optimum, and both sensitivity-change curves flatten toward its end.The final 5,000 iterations add about 0.07 dB at the deployment site and remove about 0.02 dB at the other sites.Source: C. Adaptation versus Forgetting · p. 5

V. DISCUSSION AND LIMITATIONS

Other-site OTA data transfer to the deployment site, while deployment-site data adds more at the higher-rate operating point; finetuning gives a larger local gain than joint training. These results are bounded by protocol differences and a single held-out site whose environment and mobility are confounded.

V. DISCUSSION AND LIMITATIONS

  • Other-site data transfers at both operating points, while deployment-site data adds little at 10^-1 but substantially more at 10^-2; finetuning yields greater deployment-site sensitivity than joint training at both.The OTA-trained arms lie within 0.07 dB at the 10^-1 crossing and separate toward 10^-2; simulation-only trails them at both points.Source: §V · p. 5
  • Because the same SDRs and synchronization and demodulation chain were used across sites, OTA training may learn shared hardware impairments and residual timing or frequency errors.The shared contribution is part of the other-site increment at the deployment site, 0.12 dB at 10^-2; deployment-site data adds a further 0.29 to 0.47 dB there, and controlled ablations are needed to separate effects fully.Source: §V · p. 5
  • The adaptation protocols differ in training budget, batch size, and data exposure, so the experiment compares complete protocols rather than isolating one cause.Deployment-site measurements are 3.5% of the all-site corpus and are presented about 160 times during finetuning versus about 45 times during joint training over the full budgets.Source: §V · p. 5
  • The single held-out site is both the only industrial site and the only entirely vehicular one, confounding site identity with environment and mobility.The authors identify leave-one-site-out evaluation as the natural next step.Source: §V · p. 6
  • Random rather than route- or time-blocked splits may let deployment-site finetuning and test measurements share drive routes, making some gain location-specific rather than site-wide.Source: §V · p. 6
  • Quantitative magnitudes are specific to one Hybrid CM DeepRx architecture, this campaign’s hardware, a SISO 5.88 GHz 16QAM link, and the evaluated data and protocols.No deployment-only model was trained from scratch; a comparable deployment-site corpus could yield a model that provides an upper bound for transfer and finetuning.Source: §V · p. 6

VI. CONCLUSION

The eight-site study finds that OTA data from seven source sites transfer to a distinct held-out deployment site, with deployment-site data adding more at the higher-rate operating point. Joint training preserves aggregate other-site sensitivity, while finetuning yields a larger local gain with some spatial forgetting.

VI. CONCLUSION

  • Finetuning gives the larger deployment-site gain but costs about 0.1 dB of other-site sensitivity at the higher-rate point, while remaining about 0.4 dB ahead of simulation-only on the other-site test set.These findings are based on a single held-out site.Source: §VI · p. 6
  • Under the tested protocols, joint training is preferable when aggregate other-site retention is the priority, while finetuning suits cases where deployment-site sensitivity justifies a small loss elsewhere.Source: §VI · p. 6
Loading 2610.02577v2…