Source-linked AI summary
End-to-End Optical Semantic Communication over a Nonlinear WDM Fiber Link
Hussein Jammal, Andrea Bianco, Cristina Rottondi
TL;DR
Optical applications may need task-relevant information rather than bit-exact recovery, while nonlinear WDM links complicate end-to-end optimization. This work optimizes a semantic image transmitter and receiver over a black-box Manakov-SSFM channel, reporting near-99% accuracy through 800 km with fewer symbols than LDPC-coded JPEG.
Problem
Optical applications may need task-relevant information rather than exact source reproduction, while conventional optical systems target reliable bitstream recovery and nonlinear WDM channel models challenge end-to-end optimization.
Source: §I · p. 1Method
The system jointly performs image classification and reconstruction using a semantic transmitter and receiver optimized over a black-box, non-differentiable Manakov-SSFM channel without explicit source compression or channel coding.
Source: §I · p. 1Results
98.92–99.31% classification accuracy at 0 dBm across 150–800 km and all tested QAM orders, with 128 versus 466/311/233 symbols compared with LDPC-coded JPEG at the tested modulation orders.
Source: B. Reach at nominal launch power · p. 5, C. Transmission cost · p. 6Takeaways & Limitations
In the evaluated nonlinear WDM links, semantic communication extends usable reach while reducing transmission cost relative to LDPC-coded JPEG.
Source: §V · p. 6
Abstract
from arXiv · showhide
Emerging optical-network applications increasingly use received data for inference and control rather than exact source reproduction, creating an opportunity to trade bit-level fidelity for greater transmission reach and efficiency. We propose an end-to-end optical semantic communication system for joint image classification and reconstruction over a nonlinear wavelength-division multiplexed (WDM) fiber channel. The system maps each image directly into a fixed-length sequence of channel symbols that preserves task-relevant information, without explicit source compression or channel coding. Experiments on the MNIST dataset cover launch powers from -9 to +3 dBm, fiber lengths up to 800 km, and 16-, 64-, and 256-Quadrature Amplitude Modulation (QAM) formats. At 0 dBm, classification accuracy remains between 98.92% and 99.31% across all tested link lengths and modulation orders, while requiring fewer transmitted symbols than a Low-Density Parity-Check (LDPC)-coded JPEG baseline at every tested modulation order. These results show that semantic communication can simultaneously extend optical reach and reduce transmission resources by conveying only task-relevant information.
I. INTRODUCTION
Many networked applications need task-sufficient information rather than exact source reproduction, but conventional optical systems target bitstream recovery. The paper addresses the challenge of optimizing optical semantic communication over nonlinear WDM links without channel derivatives.
I. INTRODUCTION
- Task-oriented applications need information sufficient for their task, while conventional optical systems are designed for reliable recovery of the transmitted bitstream.This mismatch becomes relevant as applications use optical backbone, data-center, and access networks.Source: §I · p. 1
- Nonlinear WDM links combine dispersion memory, accumulated ASE noise, and waveform- and launch-power-dependent Kerr distortion, complicating differentiable channel modeling.Differentiable Gaussian approximations may miss temporal correlations and nonlinear symbol interactions, whereas SSFM captures these effects but does not readily provide channel derivatives.Source: §I · p. 1
- The paper proposes joint image classification and reconstruction over a black-box, non-differentiable Manakov-SSFM channel, without explicit source compression or channel coding.It considers learned and fixed QAM constellations and a dual-head receiver.Source: §I · p. 1
- Training alternates gradient-based receiver updates with policy-gradient transmitter updates after AWGN-surrogate pretraining, and evaluation spans launch powers, link lengths, and QAM orders.The paper compares against JPEG pipelines with and without LDPC coding; the semantic system preserves task performance across tested conditions and uses fewer symbols than LDPC-coded JPEG.Source: §I · p. 2
II. SYSTEM MODEL AND END-TO-END OPTIMIZATION
The system jointly optimizes a semantic transmitter, receiver, and optionally the constellation for classification and reconstruction over a channel. Its transmitter maps each image into a normalized block of constellation symbols.
A. Problem Formulation
- The formulation maps a labeled image through a transmitter, WDM fiber channel, and receiver, which outputs a reconstructed image and class logits.The channel representation includes link length and launch-power, amplifier-noise, and neighboring-channel realizations.Source: A. Problem Formulation · p. 2
- The objective minimizes a semantic loss for classification and reconstruction, with constellation optimization included only in the learned-constellation variant.The forward system provides task-level feedback through the common loss.Source: A. Problem Formulation · p. 2, B. Semantic Transmitter · p. 2
B. Semantic Transmitter
- A convolutional encoder and logit head map each image to per-symbol categorical distributions, from which constellation points are sampled.The constellation begins as unit-average-energy Gray-labeled square QAM; its geometry is jointly optimized in the learned variant and retained in the fixed variant.Source: B. Semantic Transmitter · p. 2
- Each transmitted block is normalized to unit average symbol energy, separating launch-power control from constellation geometry.This prevents reducing the semantic loss solely by increasing block energy.Source: B. Semantic Transmitter · p. 2
C. Optical Channel Model
The optical link models a WDM signal propagated through amplified fiber and processed by coherent DSP. The proposed system and JPEG baselines differ in their treatment of carrier phase.
C. Optical Channel Model
- The channel of interest is multiplexed with power-matched neighboring channels carrying independent 16-QAM symbols, with joint propagation including inter-channel nonlinear interference.Keeping neighbor modulation fixed across the channel-of-interest QAM sweep makes the neighboring-channel environment consistent across modulation orders.Source: C. Optical Channel Model · p. 2
- Manakov propagation is numerically solved by SSFM with EDFA span compensation and ASE noise, followed by coherent DSP including dispersion compensation and matched filtering.The model includes Kerr interactions but sets the orthogonal polarization component to zero and does not model polarization-mode dispersion.Source: C. Optical Channel Model · p. 2
- The semantic receiver handles residual nonlinear phase rotation without carrier-phase recovery, while JPEG baselines receive ideal blockwise maximum-likelihood phase recovery.The paper describes this baseline choice as an optimistic performance bound.Source: C. Optical Channel Model · p. 3
D. Semantic Receiver
The receiver combines soft max-log demodulation with a neural semantic decoder. The decoder maps the resulting soft information to separate image-reconstruction and classification outputs.
D. Semantic Receiver
- The receiver rescales the normalized block using the known transmit-side factor before soft max-log demodulation against the shared constellation.The demodulator produces bit-position log-likelihood ratios for each recovered symbol.Source: D. Semantic Receiver · p. 3
- The LLRs are stacked across symbols and bit positions and passed to the semantic decoder without hard decisions or channel decoding.The demodulator's bit-position subsets correspond to constellation points labeled 0 and 1.Source: D. Semantic Receiver · p. 3
- The neural decoder uses independent heads to reconstruct the image and produce class logits, with both outputs trained jointly through the semantic loss.The reconstruction path upsamples a spatial feature volume, while the classification path maps the LLR vector through fully connected layers.Source: D. Semantic Receiver · p. 3
E. Training Procedure and Alternating Optimization
Training addresses a non-differentiable fiber channel and categorical symbol sampling through AWGN pretraining followed by alternating optimization over the SSFM channel.
E. Training Procedure and Alternating Optimization
- The SSFM channel has no available Jacobian for learning, and categorical sampling has no pathwise derivative with respect to the transmitter parameters.Starting from random weights also creates a cold-start problem: an unstructured symbol mapping yields noisy policy-gradient rewards.Source: E. Training Procedure and Alternating… · p. 3
- Differentiable AWGN pretraining initializes the models before optimization switches to the SSFM channel.Pretraining addresses the cold start caused by noisy reward signals from an initially unstructured mapping.Source: E. Training Procedure and Alternating… · p. 3
- The training variables are transmitter parameters θ, receiver parameters ϕ, and constellation C only for the learned-constellation variant.After pretraining, receiver and transmitter optimization alternate over the SSFM channel.Source: E. Training Procedure and Alternating… · p. 3
1) Semantic Loss:
The semantic loss combines classification and image-reconstruction objectives, while training alternates receiver gradient updates with transmitter policy-gradient updates over the SSFM channel.
1) Semantic Loss:
- The semantic loss combines crossentropy for classification with MSE and SSIM for image reconstruction.MSE penalizes pixel-wise error, while SSIM preserves local structure that MSE alone may smooth.Source: 1) Semantic Loss · p. 3
- Differentiable AWGN pretraining initializes the transmitter, receiver, and learned constellation before black-box optimization.A hard straight-through Gumbel Softmax estimator selects constellation points in the forward pass and differentiates through a continuous relaxation in the backward pass.Source: 3) Phase 2: Alternating Black-Box… · p. 3
- Alternating SSFM optimization updates the receiver by gradients, then updates the transmitter with policy-gradient steps.During receiver updates, transmitter parameters are frozen and symbols are selected deterministically; no gradient passes through the SSFM channel or into θ.Source: 3) Phase 2: Alternating Black-Box… · pp. 3–4
- Transmitter updates use reward R = −Lsem after SSFM propagation, with an exponential-moving-average baseline reducing estimator variance.For efficient per-sample rewards, the structural term is approximated by 1 − exp(−MSE); true SSIM is retained for receiver training and evaluation.Source: 3) Phase 2: Alternating Black-Box… · p. 4
III. EXPERIMENTAL SETTINGS
Experiments use MNIST to evaluate fixed-length semantic transmissions across QAM orders, launch powers, and optical-channel conditions against JPEG pipelines.
III. EXPERIMENTAL SETTINGS
- Experiments use 60,000 MNIST training images and 10,000 test images, with pixel intensities normalized to [0, 1].The same test set is used across all experiments.Source: §III · p. 4
- Each image uses a fixed block of 128 symbols, with QAM orders M ∈ {16, 64, 256}.The channel of interest is placed at the center of the WDM grid for both constellation variants.Source: §III · p. 4
- Training uses 100 outer iterations, each with 50 receiver-gradient and 50 transmitter policy-gradient steps, after 10 AWGN-pretraining epochs at 20 dB SNR.Launch power is sampled uniformly from [−3, +3] dBm during fine-tuning; a separate model is trained for each link length, modulation order, and constellation variant.Source: §III · p. 4
- JPEG baselines use quality factor Q = 25, with either direct bitstream transmission or rate-1/2 LDPC coding, and classify recovered images using the same ResNet-18.A channel-free control isolates source-compression effects; failed JPEG decoding counts as a classification error and receives zero SSIM.Source: §III · p. 4
- Evaluation tests launch powers from −9 to +3 dBm, using classification accuracy for task performance and SSIM for reconstruction quality.Table I lists the waveform, WDM, and fiber parameters.Source: §III · p. 4, Table I · p. 4
A. Robustness to launch power and reach
At 16-QAM, semantic transmission maintains classification accuracy across launch powers and link lengths, while JPEG baselines are vulnerable to decoding failure; reconstruction fidelity shows a different trade-off.
A. Robustness to launch power and reach
- Accuracy varies by at most 0.6 percentage points from −9 to +1 dBm at any link length, with a loss below one point at +3 dBm.The system was trained with launch powers from [−3, +3] dBm, so stability at −9 and −6 dBm is out-of-distribution; JPEG accuracy saturates near 85% when decoding succeeds.Source: A. Robustness to launch power and reach · p. 4, Figure 2 · p. 5
- At +3 dBm, LDPC-coded JPEG accuracy falls to 73.4%, 28.8%, and 3.1% at 600, 700, and 800 km, respectively.At 800 km, uncoded JPEG reaches zero accuracy by −1 dBm; residual bit errors can invalidate the entropy-coded payload.Source: A. Robustness to launch power and reach · p. 5
- The semantic system's SSIM has an operating-power optimum between −1 and 0 dBm and remains above 0.66 throughout the sweep.The optimum balances low-power ASE noise and high-power Kerr nonlinearity.Source: A. Robustness to launch power and reach · p. 5
- JPEG SSIM is near 0.869 when decoding succeeds and near zero when it fails, while semantic-system fidelity degrades gradually beyond the baseline failure point.The learned and fixed constellations perform almost identically.Source: A. Robustness to launch power and reach · p. 5
B. Reach at nominal launch power
At 0 dBm, semantic communication maintains near-99% classification accuracy across link lengths and QAM orders, while JPEG baselines lose performance as links lengthen or QAM order increases.
B. Reach at nominal launch power
- 98.92–99.31% accuracy is maintained from 150 to 800 km at every QAM order, with at most a 0.23 percentage point difference between learned and fixed constellations.These results are at 0 dBm.Source: B. Reach at nominal launch power · p. 5
- Approximately 85% accuracy persists through 800 km with 16-QAM for LDPC-coded JPEG, but falls to 11.02% at 800 km with 64-QAM and 0.67% at 500 km with 256-QAM.The passage attributes the decline to reduced noise margin at fixed launch power; FEC extends reach but does not remove QAM-order dependence.Source: B. Reach at nominal launch power · p. 6
- Semantic-system SSIM remains between 0.808 and 0.851, while JPEG-baseline SSIM drops sharply once decoding becomes unreliable.Source: B. Reach at nominal launch power · p. 6
C. Transmission cost
The semantic system uses fewer symbols and less transmission time than LDPC-coded JPEG at every tested QAM order, while keeping its image transmission time fixed across modulation orders.
C. Transmission cost
- At 32 GBaud, transmission time is symbol count divided by symbol rate, so the semantic system's 128-symbol image payload takes 4.0 ns regardless of QAM order.The JPEG payloads instead have symbol counts and transmission times that decrease as QAM order grows.Source: C. Transmission cost · p. 6
V. CONCLUSION
- The system transmits each image using 128 symbols in a fixed 4.0 ns, reducing transmission cost versus LDPC-coded JPEG at every tested modulation order while maintaining near-99% accuracy through 800 km.Source: §V · p. 6
C. Transmission cost
- 128 versus 466/311/233 symbols and 4.0 ns versus 14.6/9.7/7.3 ns: the semantic system uses less transmission time than LDPC-coded JPEG at 16/64/256-QAM.The table covers per-image transmission cost at 16/64/256-QAM.Source: Table II · p. 6, C. Transmission cost · p. 6