PaperlayerView exact revision on arXiv

Source-linked AI summary

Semantic Split Inference for Remote Modulation Recognition

Samer Lahoud, Kinda Khawam

arXiv:2610.07682v1cs.NI

TL;DR

Remote AMR must balance sensing-node complexity, reporting cost, and recognition accuracy. The paper trains a split classifier to send a semantic report over a noisy link, with split depth and report length chosen independently. Intermediate splits provide the best accuracy–cost trade-off, and an eight-use report retains 90% of local-inference accuracy.

  • Problem

    Remote AMR must balance uplink resources against the cost of placing the complete classifier on every sensing node.

    Source: §I · p. 1

  • Method

    The method splits an AMR network between sensing node and edge server, independently varies split depth and report length, and trains through the noisy reporting channel.

    Source: §I · p. 1, §III · p. 3

  • Results

    An eight-use report retained 90% of local-inference accuracy, and at matched reporting rate the intermediate split outperformed both basic placements and a state-of-the-art collaborative scheme.

    Source: §V · p. 5

  • Takeaways & Limitations

    Intermediate splits kept edge-side capacity to process noisy reports while avoiding the highest sensing-node cost, yielding the best accuracy–cost trade-off in the evaluated setting.

    Source: §V · p. 5

Abstract

from arXiv · show

Remote automatic modulation recognition balances sensing-node complexity, reporting cost and accuracy. To address this trade-off, we propose channel-aware semantic split inference: a sensing node sends a semantic report over a noisy link and the edge server completes recognition. In our model, split depth and report length are independent design variables, with end-to-end training through the channel. We compare the resulting design against basic split placements, which run inference at the edge server or at the sensing node, and against a state-of-the-art collaborative scheme. We assess sensing-node model size, computation, latency and energy against recognition accuracy. We show that intermediate splits give the best accuracy-cost trade-off.

I. INTRODUCTION

Remote AMR must balance uplink resources against the cost of running recognition on sensing nodes. The paper proposes channel-aware semantic split inference and evaluates intermediate splits against basic placements and a collaborative scheme.

I. INTRODUCTION

  • Forwarding raw I/Q samples uses substantial uplink resources, whereas local classification requires the full model at every sensing node.These alternatives motivate distributing inference between the node and edge server.Source: §I · p. 1
  • Channel-aware semantic split inference sends a compact, classification-oriented report over a noisy channel for the edge server to complete recognition.The report is formed from intermediate network features.Source: §I · p. 1
  • Split depth and report length are independent design variables: depth trades node-side processing against representation quality, while length sets channel uses and affects accuracy.Shallow splits reduce the node model but produce less-processed representations; deeper splits incur greater local cost.Source: §I · p. 1
  • At a matched reporting rate, intermediate splits provide the best accuracy–cost trade-off against basic split placements and a state-of-the-art collaborative scheme.The evaluation includes node model size, computation, measured latency and energy, and accuracy.Source: §I · p. 1

II. RELATED WORK AND POSITIONING

Prior AMR work emphasizes recognition accuracy and device-side efficiency, while remote recognition introduces the additional question of what each sensing node should transmit. This paper positions independently chosen split depth and report length over a noisy link as the distinction from nearby split-AMR schemes.

II. RELATED WORK AND POSITIONING

  • Deep-learning AMR classifiers learn discriminative temporal representations from raw I/Q, but the reviewed literature assumes observation and classifier share a device.Hardware-aware recognizers address the cost of accuracy on constrained platforms.Source: §II · p. 1
  • Federated learning exchanges model parameters and keeps observations local, but still requires the complete inference model on every node.Split computing instead partitions a trained network between device and server.Source: §II · p. 1
  • C-AMC compresses node features before edge classification, while SplitAMC forwards a full intermediate feature and requires at least eight times raw-I/Q forwarding's channel uses.SplitAMC is also the uncompressed baseline against which C-AMC is reported.Source: §II · pp. 1–2
  • Nearby split-AMR schemes fix report formation or omit a report-length variable, so split depth and reporting rate cannot be varied independently.General split-computing links are usually modeled as error-free bit pipes.Source: §II · pp. 1–2
  • This approach optimizes the transmitted representation for classification through a stochastic link, with both sensing and reporting channels affecting remote-AMR accuracy.The report length can be divided between semantic width and token count.Source: §II · p. 2

III. SYSTEM MODEL AND SEMANTIC SPLIT INFERENCE

The system divides a residual AMR network between a sensing node and edge server, transmitting a power-normalized semantic report over a noisy reporting link. Split depth controls node-side model cost, while report dimensions control channel use; the modules are trained end to end through the reporting channel.

a) Sensing and reporting model

  • The sensing node observes N complex-baseband I/Q samples for an unknown modulation class, with sensing noise, carrier-frequency offset, and phase offset represented in the signal model.The real-valued network input stacks the in-phase and quadrature components, so raw forwarding takes 2N real channel uses.Source: §III · p. 2
  • A split at depth k assigns the first k residual stages to the node; its transmit interface forms a frame-normalized report of Lz real channel uses for edge-side completion.The report is transmitted at unit average power per use; the edge receive interface reconstructs features for the remaining stages and classification head.Source: §III · p. 2
  • The reporting link is evaluated with AWGN and flat Rayleigh block fading, while path loss, shadowing, and co-channel interference are absorbed into the effective reporting SNR under stated assumptions.For interference, the model assumes an independent interferer and no edge-side interference cancellation.Source: §III · p. 2

c) Basic split placements

  • Split depth k sets node model size and computation, while Lz sets reporting cost; the compared extremes use a barely processed report at k = 1 or the full backbone at the node at k = K.The design trades these costs against recognition accuracy.Source: §III · pp. 2–3

b) Backbone and semantic bottleneck

  • The residual CNN has K stages; a transmit projection sets semantic width nc and adaptive pooling sets token count T, with report length Lz = ncT.T = 1 gives a global representation, whereas T > 1 preserves temporal feature tokens; the edge interface restores the feature tensor dimensions.Source: §III · pp. 2–3

d) Training

  • The encoder, transmit and receive interfaces, and classifier are optimized end to end through the stochastic reporting channel using cross-entropy.SNR-aware training redraws reporting SNR per mini-batch; random phase rotation and circular time shifting augment frames while preserving class and power.Source: §III · p. 3

IV. RESULTS

The intermediate split offers the strongest accuracy–cost balance against basic placements and C-AMC. Accuracy also depends on report length, its width–time allocation, and reporting- and sensing-link conditions.

Effect of split depth

  • Accuracy rises by about 12.5 points from k = 1 to k = 3, then largely plateaus at k = 4 before falling by 6 points at k = 5 for both tested compression ratios.The k = 3 to k = 4 gains are 0.7 points at ρ = 128 and 1.6 points at ρ = 64; the latter split-depth comparison is within one SD.Source: §IV · p. 3, Figure 2 · p. 3
  • On an NVIDIA Jetson AGX Orin at 30 W, k = 3 accounts for 90% of FLOPs but 58% of inference time, consuming 6.8 mJ—about 61% of the 11.1 mJ local classifier.Source: §IV · p. 3

Effect of semantic report length

  • At γr = 20 dB, k = 3 accuracy is 0.83 at ρ = 256, stays above 0.80 through ρ = 512, and falls to 0.61 at ρ = 1024.When γr falls from 20 to 0 dB, accuracy drops by 41 points at ρ = 512, compared with 8 points at ρ = 32.Source: §IV · p. 4, Figure 3 · p. 4

Allocation of a fixed report length

  • For fixed Lz ≤ 16, the spread across report-width and temporal-resolution allocations is 13 to 14 points, but it is 2.7 points at Lz = 32.Configurations with nc ≥ 4 are within 1.3 points of the best at the same budget when Lz ≤ 16; nc = 2 never exceeds 0.75.Source: §IV · p. 4, Table II · p. 4

Flat fading on the reporting link

  • Under flat Rayleigh fading, the penalty relative to AWGN is 1.2 to 3.5 points at γr = 20 dB and reaches 10.7 points at γr = 0 dB.The accuracy loss from γr = 20 to 0 dB is 1.9 times larger under fading at ρ = 64 and 2.1 times larger at ρ = 32.Source: §IV · p. 4, Table III · p. 4

Comparison with basic placements and the state of the art

  • At Lz = 32, the k = 3 intermediate split beats both basic placements: k = 1 trails by 12.6 points, while k = 5 trails by 4.5 points despite 68% more node parameters and 66% more time.The comparison uses the matched report length and conditions specified for Table IV.Source: §IV · pp. 4–5, Table IV · p. 5
  • C-AMC uses 0.02 M node parameters versus 0.16 M for the proposed split but trails it by 13.3 points at the matched reporting rate.C-AMC's total size is similar to the proposed model, but 93% of its parameters sit at the edge server.Source: §IV · pp. 4–5, Table IV · p. 5

Sensing SNR and class-set size

  • The k = 3 split retains about 90% of local accuracy down to γs = 5 dB; with 24 classes, split-depth rankings stay unchanged, while with 8 classes depth matters only on a poor link.At γs = 0 dB, both the split and local models are close to chance.Source: §IV · p. 5, Table V · p. 5

V. CONCLUSION

The conclusion identifies intermediate splits as the best trade-off among sensing-node complexity, reporting cost, and recognition accuracy. At matched reporting rates, they outperform both basic placements and a state-of-the-art collaborative scheme.

V. CONCLUSION

  • Intermediate splits give the best trade-off and outperform both basic placements and a state-of-the-art collaborative scheme at matched reporting rates.The design decouples network split depth from semantic report length.Source: §V · p. 5
  • An eight-use report retains 90% of local-inference accuracy.Source: §V · p. 5
  • Running the whole backbone at the sensing node has the highest node cost and is still less accurate, because no edge-side capacity remains to process the noisy report.Source: §V · p. 5
  • The intermediate split keeps its advantage with 24 classes and retains about 90% of local accuracy at lower sensing SNR.Source: §V · p. 5
  • Increasing observation bandwidth raises encoder computation, while adaptive pooling keeps report length fixed by construction.Source: §V · p. 5
  • Multi-antenna observation, faster channel variation, frequency-selective reporting links, and class-set mismatch remain for future work.Source: §V · p. 5
Loading 2610.07682v1…