PaperlayerView exact revision on arXiv

Source-linked AI summary

Significance-Driven Semantic Communication

Christian McDowell, Andrea Panebianco, Sirin Chakraborty, Yin Sun

arXiv:2608.28441v1cs.IT

TL;DR

The paper addresses the lack of rigorous measures for the task-specific value of individual observations and the resulting separation of semantic encoding from resource allocation. It introduces a per-sample significance measure, Meta-VIB, and Q-Maximization within a joint cross-layer design. On real-world pedestrian safety data, the design achieves up to 1000 times and 40 times gains in semantic spectrum efficiency at average SNRs of 0 dB and 5 dB, respectively.

  • Problem

    Existing communication systems lack a rigorous framework for quantifying individual samples’ task-specific value and often treat physical-layer design and MAC scheduling separately.

  • Method

    The paper jointly optimizes semantic encoding and inference with MAC allocation using per-sample significance, Meta-VIB, and Q-Maximization.

  • Results

    Up to 1000 times and 40 times gains in semantic spectrum efficiency are reported at average SNRs of 0 dB and 5 dB, respectively, versus cross-layer baselines.

  • Takeaways & Limitations

    Per-sample significance provides the basis for a cross-layer semantic communication design that prioritizes task-relevant value in encoding and channel allocation.

Abstract

from arXiv · show

In this paper, we study a significance-driven cross- layer semantic communication design problem. Based on sta- tistical decision theory, we introduce an information-theoretic measure of per-sample data significance that quantifies the task-specific value of each individual observation. Using this metric, we formulate a cross-layer optimization problem that simultaneously optimizes (i) physical-layer semantic encoding and inference and (ii) MAC-layer resource allocation, with the objective of maximizing semantic spectrum efficiency, defined as the semantic value delivered per unit bandwidth per unit time. At the physical layer, we develop Meta-Learning Variational Information Bottleneck (Meta-VIB), a new semantic transceiver that employs a meta-learned hypernetwork to compress high- dimensional observations into semantically significant latents, enabling instantaneous adaptation to dynamic channel conditions and varying symbol budgets without online retraining. At the MAC layer, we model channel allocation as a Multi-Action Restless Multi-Armed Bandit (MA-RMAB) and adopt the Q- Maximization algorithm, which dynamically allocates channel resources to sensors based on their semantic value of information. Experimental results on a real-world pedestrian safety dataset demonstrate that our joint design achieves substantial gains in semantic spectrum efficiency over baselines, reaching up to 1000 times gain at an average SNR of 0 dB and 40 times gain at an average SNR of 5 dB.

I. INTRODUCTION

The paper develops a significance-driven semantic communication design because existing systems lack rigorous per-sample value measures and often prioritize freshness or data type over task-specific relevance. It jointly optimizes physical-layer semantic encoding and inference with MAC-layer resource allocation to maximize semantic spectrum efficiency.

  • Motivation: Existing communication systems often use freshness or data type as indirect proxies instead of task-specific data value.This disconnect makes it harder to prioritize information relevant to downstream inference and decision-making.
  • Contribution: The proposed cross-layer design jointly optimizes physical-layer semantic encoding and inference with MAC-layer channel allocation.Its objective is semantic spectrum efficiency: task-relevant value delivered per unit bandwidth per unit time.
  • Contribution: The paper introduces an information-theoretic per-sample significance measure based on statistical decision theory and a task-specific loss function.The measure quantifies the value of each individual observation for the downstream task.
  • Physical-layer design: Meta-VIB adapts semantic encoding across time-varying channels and codelength constraints without online retraining.The transceiver uses FiLM conditioning, codelength truncation, an information-concentrating regularizer, and a specialized training procedure.
  • MAC-layer design: Q-Maximization dynamically allocates MAC-layer channel resources across sensors in the resulting cross-layer design.The scheduling problem is modeled as a Multi-Action, Restless Multi-Armed Bandit over a high-dimensional continuous state space.
  • Evaluation: 1000 times gain in semantic spectrum efficiency is achieved at an average SNR of 0 dB, while 40 times gain is achieved at 5 dB.These gains are reported for the (Meta-VIB, Q-Maximization) design versus cross-layer baselines on a pedestrian safety dataset.

B. Information-Theoretic Interpretation

The paper gives per-sample semantic value a decision-theoretic and information-theoretic interpretation by defining loss-based entropy, divergence, and mutual information. It shows that an observation’s conditional value equals an L-divergence between posterior and prior distributions, while logarithmic loss recovers Shannon mutual information.

  • Information measures: L-entropy is the minimum expected loss for predicting Y without side information, while L-conditional entropy is the average posterior Bayes risk.The framework defines these quantities for a task-dependent loss function L.
  • Information measures: L-divergence measures the excess expected loss from using a Bayes action optimized for QY instead of the true distribution PY.It is non-negative and reduces to Kullback–Leibler divergence under logarithmic loss.
  • Information-theoretic interpretation: Under logarithmic loss, L-mutual information becomes symmetric Shannon mutual information; for general losses, it need not be symmetric.Thus the framework is loss-dependent rather than universally equivalent to ordinary mutual information.
  • Information-theoretic interpretation: Proposition 3.1 identifies the conditional value vL,Y(z) exactly with the L-divergence between PY|Z=z and PY.Averaging this quantity over Z yields the corresponding L-mutual information.
  • Semantic significance: The framework emphasizes instantaneous per-sample significance vL,Y(z), rather than only average value of information.This conditional quantity captures the semantic value of a specific observation for significance-aware communication and scheduling.

IV. SIGNIFICANCE-DRIVEN SEMANTIC COMMUNICATION: SYSTEM MODEL AND PROBLEM FORMULATION

The section models significance-driven semantic communication from sensing through decoding and scheduling, then formulates a joint optimization targeting semantic spectrum efficiency.

  • A. System Model: The network comprises edge sensors that extract historical semantic features from raw frames for downstream task inference.Each sensor forms X_n,t from a configurable history window intended to capture task-relevant temporal context.
  • A. System Model: The latent codewords traverse an orthogonal multiple-access block-fading channel with additive complex Gaussian noise and shared symbol constraints.The channel state is represented by h_n,t, while ζ_n,t denotes SNR and total allocation is limited by available bandwidth and time.
  • A. System Model: When a sensor is unscheduled, the receiver reuses its freshest received latent, while AoI resets after transmission and increments otherwise.The decoder uses the time-stamped latent together with its age to support inference from stale information.
  • A. System Model: The decoder maps received time-stamped latents to task estimates under a task-dependent loss, while the scheduler maps receiver information to codelength allocations.These components define the inference objective and admissible scheduling policy.
  • B. Information Structure: The encoder, decoder, and scheduler operate on distinct information sets reflecting their sensor-side and receiver-side roles.The encoder uses local source, channel, and codelength information; the decoder uses received signal, channel, and codelength; the scheduler uses aggregate sample and SNR states.
  • C. Cross-Layer Optimization Problem and Semantic Spectrum Efficiency: The cross-layer problem jointly optimizes physical-layer encoder-decoder parameters and the MAC-layer scheduling policy.The objective accumulates discounted semantic significance across the network, with mutual information and channel capacity entering the resource trade-off.
  • C. Cross-Layer Optimization Problem and Semantic Spectrum Efficiency: Semantic spectrum efficiency measures semantic significance delivered per unit bandwidth per unit time rather than transmitted bits per unit spectrum.The metric evaluates how effectively spectrum delivers task-relevant information for downstream inference.

V. SIGNIFICANCE-DRIVEN SEMANTIC COMMUNICATION:

The paper decomposes the cross-layer problem into physical-layer semantic encoding and MAC scheduling, then uses Meta-VIB to handle changing operating conditions without retraining.

  • A. Problem Decomposition: Problem (20) jointly optimizes physical-layer semantic encoding and inference with MAC-layer codelength allocation through the shared codelength process.The allocation determines both each sensor’s symbol budget and its scheduling-state AoI.
  • A. Problem Decomposition: The physical-layer encoder-decoder is designed to maximize expected semantic significance of received codewords across possible AoI values.The design remains useful over AoI values from 1 through δ_max and must handle changing channel and codelength conditions.
  • A. Problem Decomposition: Solving a distinct encoder-decoder for every dynamic channel and codelength realization would be computationally prohibitive without adaptation.The paper addresses this requirement with Meta-VIB rather than online retraining.
  • A. Problem Decomposition: The MAC layer substitutes optimized physical-layer significance into a scheduling problem over sensor states and codelength actions.The expected delivered significance is represented by g_n under each scheduling state and allocation.
  • A. Problem Decomposition: The scheduling problem is a Multi-Action Restless Multi-Armed Bandit with a high-dimensional continuous state space.Multi-discrete actions invalidate classical binary-action Whittle-index assumptions, while unknown dynamics create a curse-of-dimensionality challenge.
  • A. Problem Decomposition: Q-Maximization is used to address the scheduling challenges created by multi-action decisions and data-driven high-dimensional states.The supplied passages identify this algorithm as the proposed solution for the MAC layer.
  • B. Physical-Layer Design: Meta-VIB: Meta-VIB adapts to time-varying channel conditions and codelengths without online retraining.Its objective is derived as a loss-based information bottleneck, with an MMD penalty reducing mismatch between aggregated posterior and Gaussian prior.
  • 1) Meta-VIB Formulation:: The Meta-VIB objective replaces standard VIB’s KL-based task relaxation with an L_n-conditional-entropy relaxation aligned with task loss across AoI values.The information-rate term is bounded using a variational Gaussian prior and KL divergence.

2) FiLM-based Lightweight Meta-VIB Design.:

Meta-VIB uses FiLM conditioning through a lightweight hypernetwork to adapt one semantic transceiver across channel conditions, codelengths, and regularization settings without retraining. Its design combines flexible latent truncation, information concentration, staged training, and online dual optimization.

  • FiLM Conditioning: Meta-VIB uses a single FiLM-conditioned architecture that generalizes across SNR, codelength, and β settings without retraining.The hypernetwork modulates encoder and decoder features based on the operating point.
  • Backbone Architecture: The encoder uses a BiLSTM backbone, while the decoder uses residual MLP blocks with GELU activations.
  • FiLM Conditioning: FiLM modulation is generated by a lightweight MLP hypernetwork without changing the backbone weights.The hypernetwork outputs modulation parameters in one forward pass.
  • Latent Truncation: The information-concentrating regularizer orders posterior variances so prefix truncation discards the least informative latent suffix first.
  • Training Procedure: Meta-VIB is trained through pretraining, hypernetwork training, and joint fine-tuning across discrete β and codelength settings.The three phases occupy 40%, 45%, and 15% of total training epochs, respectively.
  • Online Dual Optimization: At deployment, the optimal β is computed online for each codelength and SNR realization using golden-section search over log β.The search converges in 25 iterations.

C. MAC-Layer Design: Q-Maximization Algorithm

The MAC layer formulates channel allocation as an MA-RMAB and uses dual decomposition to obtain per-sensor decision processes. Q-Maximization then allocates discrete codelengths under a shared spectrum budget using learned Q-values and LP-priority tie-breaking.

  • MA-RMAB Formulation: Lagrangian relaxation converts the hard per-slot spectrum constraint into independent per-sensor MDPs with reward gn(ωn,t, ηn,t) − ληn,t.
  • Offline Training: A single offline actor-critic network learns per-sensor Q-functions over a continuous range of dual variables λ.This supports adaptation when the sensor count or available bandwidth changes.
  • Online Scheduling: At deployment, bisection computes λ* for the current configuration before Q-values determine each sensor’s codelength allocation.
  • Online Scheduling: Q-Maximization solves a Multiple-Choice Knapsack Problem subject to the global spectrum budget and discrete per-sensor codelength choices.
  • Algorithmic Properties: Q-Maximization retains O(NWK) complexity while adding LP-priority tie-breaking and is asymptotically optimal for finite-state, finite-action MA-RMABs.

VI. DATA-DRIVEN EVALUATION

The evaluation uses synchronized live traffic-camera feeds and compares Meta-VIB and Q-Maximization with physical- and cross-layer baselines. Across tested conditions, the proposed design improves semantic significance and semantic spectrum efficiency.

  • Data Collection: The dataset consists of synchronized traffic-camera feeds covering four crosswalks at a busy downtown Auburn, Alabama intersection.A Raspberry Pi 5 recorded 1080p streams at 20 fps from 06:00 to 18:00 over two weeks.
  • Baselines: Meta-VIB is compared with JSCC, DeepJSCC, VIB, Hyper-VIB, and ATROC, while scheduling baselines include SemanticGreedy, MaxAge, and Round-Robin.The physical-layer baselines use fixed 10 dB SNR and zero AoI, whereas Meta-VIB trains over varying SNR and AoI ranges.
  • Physical-Layer Evaluation: As SNR increases, per-sample semantic significance increases, and Meta-VIB consistently achieves higher semantic value than DeepJSCC.
  • Physical-Layer Evaluation: Average semantic significance increases with codelength and SNR but decreases with AoI, with Meta-VIB approaching the theoretical upper bound closely.
  • Cross-Layer Evaluation: 1000 times gain in semantic spectrum efficiency is achieved by (Meta-VIB, Q-Maximization) at average SNR E[ζn,t] = 0 dB.Q-Maximization also outperforms NGM through its LP-priority-based tie-breaking step.
  • Conclusion: The proposed significance-driven cross-layer design consistently improves semantic spectrum efficiency under varying bandwidth and SNR conditions.

APPENDIX A PROOF OF PROPOSITION 3.1

This appendix develops the Lagrangian and dual-decomposition machinery underlying MAC-layer scheduling, including per-sensor Q-functions and dual optimization. It also presents the Q-Maximization framework and supporting significance-related figures.

  • Lagrangian relaxation: The MAC problem relaxes its coupled capacity constraint into a discounted budget constraint and associates a nonnegative dual multiplier with that budget.The relaxed formulation enables decomposition across sensors.
  • Dual decomposition: The penalized reward decomposes across sensors, yielding independent discounted MDPs with per-slot reward gn(ωn,t, ηn,t) − ληn,t.Each sensor optimizes scheduling actions for a fixed dual price λ.
  • Per-sensor solution: The Bellman optimality equation remains valid for high-dimensional continuous scheduling states, while finite known-state problems can alternatively use classical dynamic programming.The appendix motivates learned Q-functions for the continuous-state setting.
  • Dual optimization: Bisection finds λ* by enforcing U(λ*, α) = W/(1 − γ), increasing λ when discounted resource usage exceeds the budget.The resource-usage function is nonincreasing in λ.
  • Q-Maximization: Q-Maximization belongs to the multi-action LP-priority policy class and passes the resulting per-sensor Q-values to the scheduler.Its priority rules are recovered from Q-values rather than requiring the LP solution directly.

A. Fluid LP Relaxation

The relaxed MAC problem is reformulated as a fluid linear program using discounted occupancy measures. Its optimal value provides an upper bound on the original problem’s objective for every sensor count.

  • Fluid LP formulation: The relaxed problem is reformulated as a linear program by introducing discounted occupancy measures.This produces the fluid LP representation used for the asymptotic analysis.
  • Role in analysis: The fluid LP supplies the benchmark objective later used to characterize asymptotic scheduling performance.This follows from its stated role as an upper bound in the appendix’s analysis.
  • Upper bound: Rγ,fluid, the optimal objective value of the fluid LP, upper-bounds the optimal objective value of the original problem for every N.The bound applies before taking any large-system limit.

B. Multi-Action LP-Priority Rules

The multi-action LP-priority rules rank sensor-state-action triples according to LP occupancy structure and optimal actions. Under mild technical conditions, policies obeying these rules are asymptotically optimal.

  • State categories: The LP partitions sensor-state pairs into pure-action, mixed-action, and unvisited categories according to their optimal occupancy structure.Unvisited pairs have zero occupancy for every action.
  • Priority ordering: Pure-action pairs receive highest priority, mixed-action pairs follow, and unvisited pairs receive lowest priority.Within mixed-action pairs, larger LP-optimal codelengths receive higher priority.
  • Action selection: Only actions in the LP-optimal action set are assigned for pure-action and mixed-action pairs, while arbitrary fixed actions may be used for zero-occupancy pairs.The arbitrary choice does not affect fluid-limit performance because those pairs have zero occupancy.
  • Asymptotic guarantee: As N and W grow proportionally to infinity, every policy in Π* converges in discounted significance to the fluid LP upper bound Rγ,fluid.The result holds under the mild technical conditions of.

C. Modified Dynamic Programming with LP-Priority Tie-Breaking

The modified dynamic program implements Q-Maximization under a channel budget while preserving LP-priority tie-breaking. It selects Q-maximizing actions, preferring larger codelengths when maximizers are tied.

  • Q-value implementation: Q-Maximization recovers LP-priority rules directly from per-sensor Q-values without requiring the LP-optimal solution x*.The learned Q-function is evaluated at an augmented scheduling state.
  • Action selection: For a unique Q-maximizing action, the algorithm assigns that action, including η = 0 when the sensor remains unscheduled.This corresponds to a pure-action pair.
  • Tie-breaking: For tied Q-maximizing actions, the algorithm assigns the largest codelength that fits the residual budget and otherwise tries the next largest.Actions outside the maximizing set are not assigned while a maximizing action fits.
  • Performance: Under technical conditions and proportional growth of N and W, Q-Maximization converges to the fluid LP upper bound in discounted significance.For continuous high-dimensional states, the per-sensor Q-functions are learned with a λ-conditioned actor-critic procedure.
  • Dynamic programming: The modified MCKP dynamic program changes the forward fill to retain the largest tied codelength, while leaving backtracking unchanged.This addresses arbitrary standard tie-breaking that can leave budget unused.

D. Integration with the RL Architecture

The evaluation integrates significance-aware physical-layer compression with receiver-side scheduling and examines performance across codelength, channel conditions, and information age. Meta-VIB remains advantageous under resource scarcity and stale information, while the joint design outperforms cross-layer baselines in a favorable scenario.

  • RL architecture: The λ-conditioned RL architecture learns Q-values from continuous, high-dimensional scheduling states and recovers partitions through Q-value ties with priority-based tie-breaking.Deployment passes the learned Q-values to Algorithm 3 for partition recovery.
  • Physical-layer evaluation: The advantage of significance-driven compression is most pronounced in the low-η regime and at high δ, where resource scarcity and information staleness reduce communication utility.These conditions make significance-driven compression particularly important for maintaining utility.
  • Online scheduling: The receiver evaluates trained Q-values for every available codelength and uses Algorithm 3 to produce the allocation in O(NWK) time.The scheduler operates entirely from receiver-side information after the physical layer is frozen.
  • Physical-layer evaluation: Meta-VIB consistently achieves higher significance than baselines and approaches the upper bound H_L(Y_n,t+δ) across codelength and instantaneous SNR.The average semantic value increases with both codelength and SNR in the evaluated favorable scenario.
  • Cross-layer evaluation: The Meta-VIB and Q-Maximization design outperforms cross-layer baselines as semantic spectrum efficiency decreases with W and increases with N.Fixed-codelength baselines cannot increase codeword length when instantaneous channel capacity is insufficient, while NGM lacks Q-Maximization’s asymptotic optimality.
Loading 2608.28441v1…