Adaptable Semantic Compression and Resource Allocation for Task-Oriented Communications

Motivation

• Task-oriented communication transmits meaningful raw-data content for intelligent tasks instead of requiring accurate reception of every bit.The paradigm is intended to reduce network traffic and alleviate spectrum shortage.Source: §I · p. 1

Open challenges

• Semantic compression is task-dependent because semantic information is abstract and subjective, creating a challenge for unified compression criteria.The same data may have different semantics across intelligent tasks.Source: §I · p. 1
• Semantic compression reduces transmitted data and computing overhead, while resource allocation is needed to address limited wireless resources and prioritized reliability requirements.The paper identifies resource management as necessary because compression alone does not resolve wireless-resource limitations.Source: §I · p. 2

Proposed framework

• The proposed E2E framework extracts, compresses, and transmits raw-data semantics to an edge server, which performs intelligent computing and returns task results.The system uses semantic and channel encoders at the transmitter and corresponding decoders at the receiver.Source: §II · pp. 2–3, A. E2E Task-Oriented Semantic… · p. 3

Adaptable semantic compression

• Adaptable semantic compression transmits semantic features whose importance exceeds a compression-ratio-determined threshold and discards the others.Existing semantic communication systems are represented as the special case ω0 = 0.Source: A. E2E Task-Oriented Semantic… · p. 3

Optimization objective

• The system improves success transmission probability by jointly optimizing semantic compression ratios and wireless resource allocation.Success transmission probability is mainly affected by power, bandwidth, and semantic compression ratio.Source: A. E2E Task-Oriented Semantic… · p. 4

• The success probability of tasks evaluates semantic communication by jointly reflecting transmission and semantic understanding.It controls the tradeoff between semantic transmission and semantic understanding.Source: B. Problem Formulation · p. 4
• Users may receive different service levels, with higher-priority users requiring performance protection.For example, fire-detection users in smart factories may require highly reliable, low-latency communication.Source: B. Problem Formulation · p. 4
• Only a subset of users may be selected because limited wireless resources must preserve higher-priority user performance.The selection index β_i indicates whether user i is selected.Source: B. Problem Formulation · p. 4
• The optimization jointly chooses resource allocation, compression ratios, and user selection to maximize weighted sum success probability under system constraints.Constraints cover total and minimum bandwidth, total and minimum transmit power, compression ratios, and service-level assignment.Source: B. Problem Formulation · pp. 4–5

• The section develops an end-to-end semantic communication system and uses semantic-feature importance as the basis for adaptable semantic compression.It also studies how intelligent-task performance varies with compression ratio.Source: §III · p. 5

• Deep neural networks implement a joint encoder-decoder framework for capturing and transmitting meaningful source semantics.The framework targets complex source data, including image recognition tasks.Source: A. E2E Semantic Communication Network… · p. 5
• The semantic transceiver is jointly trained end-to-end with a neural-network channel model under an AWGN channel assumption.Gradients are backpropagated from the semantic decoder output to the semantic encoder input.Source: A. E2E Semantic Communication Network… · p. 5

B. Loss Function Design

• The loss combines task loss with a weighted mutual-information upper-bound term to improve task performance while increasing achieved data rate.The task term can use cross-entropy or triplet loss, while κ weights mutual information between transmitted and received semantics.Source: B. Loss Function Design · pp. 5–6
• The mutual-information upper bound is optimized using a variational distribution qθ(y|x) implemented by neural networks.The bound exceeds the true mutual information, and its variational objective is minimized by gradient descent.Source: B. Loss Function Design · pp. 5–6
• Training uses two phases: unsupervised estimation of the achieved data rate, followed by joint-system training with the combined loss.Both phases use mini-batch gradient descent, and adaptable semantic compression is applied only during inference.Source: B. Loss Function Design · p. 6

• Semantic importance is defined as the correlation between semantic features and the intelligent task.Features irrelevant to the task represent semantic redundancy that can be further compressed.Source: C. Semantic Importance Evaluation · p. 6
• A gradient-based method measures feature importance by differentiating a semantic concept activation with respect to feature activations.The gradients are globally averaged over the feature-map width and height to obtain importance weights.Source: C. Semantic Importance Evaluation · p. 6
• The resulting weight captures the semantic importance of a feature for a semantic concept through a partial linearization of the downstream network.It is computed from successive products of network weights and activation-function gradients.Source: C. Semantic Importance Evaluation · p. 6
• Importance weights depend only on network parameters, so sender and receiver can store them as shared task-specific knowledge without transmitting feature-map indices.The paper uses the gradient-based method to calculate these weights.Source: C. Semantic Importance Evaluation · p. 6

• A closed-form task-performance function η(o) is intractable because it depends on neural-network behavior.Source: D. Intelligent Task Performance Model · p. 6
• The fitting parameters ζ are learned iteratively from sampled compression-ratio and task-performance points.Source: D. Intelligent Task Performance Model · p. 7
• Feature-map importance is used to remove unimportant maps progressively and measure corresponding task performance and compression ratios.Source: D. Intelligent Task Performance Model · p. 7
• The estimated mapping is fitted with η(o) = ζ1e^ζ2o + ζ3e^ζ4o.Source: D. Intelligent Task Performance Model · p. 7
• For equal-service-level users, all users are selected and the optimization focuses on compression ratios and resource allocation.Source: §IV · p. 7
• Because the objective is non-concave, CRRA separates the problem into two subproblems and solves them iteratively until convergence.Source: §IV · p. 7

• With a fixed resource-allocation policy, each user’s optimal semantic compression ratio is independent of the others.Source: A. Optimal Compression Ratios · p. 7
• Because o_i lies between 0 and 1, one-dimensional enumeration searches candidate ratios to obtain each user’s optimum.Source: A. Optimal Compression Ratios · p. 7
• The resource-allocation stage treats η(o_i) as a constant α_i once compression ratios are fixed.Source: B. Optimal Resource Allocation · p. 7
• Slack variables and successive convex approximation transform the resource-allocation formulation into a convex problem.Source: B. Optimal Resource Allocation · pp. 7–8
• The convex resource-allocation problem is solved iteratively, updating variables until the objective converges.Source: B. Optimal Resource Allocation · p. 8
• CRRA alternates compression-ratio enumeration with resource allocation and repeats until the objective value converges.Source: B. Optimal Resource Allocation · p. 8

• CRRAUS targets users with different task priorities and consequently different service levels.Source: §V · p. 8
• Only part of the users may be selected because wireless resources are constrained.Source: §V · p. 8
• The algorithm adaptively adjusts user selection according to wireless resources and service levels.Source: §V · p. 8

• The original problem is difficult because its objective function is non-concave and its constraints are nonconvex.Source: A. Algorithm Design · p. 8
• CRRAUS first optimizes compression ratios with fixed resources and selection, then solves user selection, and finally optimizes resource allocation.Source: A. Algorithm Design · p. 9
• Compression-ratio optimization maximizes each user’s weighted task-success probability and uses one-dimensional enumeration.Source: A. Algorithm Design · p. 9
• User selection is formulated as a 0-1 integer program and solved with branch and bound.Source: A. Algorithm Design · p. 9
• Branch and bound branches on fractional variables and updates upper and lower bounds until an integer solution is obtained.Source: A. Algorithm Design · p. 9
• With selection and compression ratios fixed, resource allocation uses SCA and the dual method, while all three subproblems are iterated to convergence.Source: A. Algorithm Design · p. 9

Convergence

• CRRA’s total success probability is monotonically non-decreasing, and its iterates converge to a point satisfying the original problem’s KKT conditions.The convergence analysis focuses on the resource-allocation subproblem, while compression ratios are solved by one-dimensional enumeration.Source: B. Convergence and Complexity Analysis · p. 9

CRRA complexity

• CRRA has total complexity O(T0KU + T0U^3.5), combining compression-ratio enumeration with resource-allocation optimization across T0 iterations.With fixed resource allocation, compression-ratio optimization costs O(KU); with fixed compression ratios, resource allocation costs O(U^3.5).Source: B. Convergence and Complexity Analysis · p. 9

CRRAUS complexity

• CRRAUS adds one-dimensional enumeration, branch-and-bound user selection, and SCA-based optimization to its computational complexity.The algorithm addresses compression ratios, user selection, and resource allocation through separate optimization components.Source: B. Convergence and Complexity Analysis · p. 10

Simulation results

• At 80% semantic compression, classification-accuracy loss is tiny above 0 dB SNR while transmitted data and delay are reduced.The comparison uses VGG and ResNet backbones and contrasts ASC with traditional communication and lower compression levels.Source: §VI · p. 11
• Optimal compression ratios vary with available bandwidth and transmit power, reflecting a trade-off between transmission efficiency and task performance.The reported optimum is 0.7 at 10 MHz and 0.8 at 20 MHz bandwidth.Source: §VI · p. 11
• CRRA outperforms FCR, FRA, and MSR in task success probability, especially with many users, while incurring additional computational complexity.The advantage is attributed to jointly choosing compression ratios and resource allocation under delay constraints.Source: §VI · p. 13

• The proposed architecture adaptively compresses semantics to reduce transmitted data and communication burden.Source: §VII · p. 13
• CRRA iteratively optimizes resource allocation and compression ratios for users with the same service levels.Source: §VII · p. 13
• CRRAUS adaptively selects users with varying service levels using a branch and bound method.Source: §VII · p. 13
• Both CRRA and CRRAUS achieve higher task success probability than benchmark algorithms in simulations.Source: §VII · p. 13

• The proof begins by using equations (6) and (7) to establish the lemma's intermediate result.Source: Appendix A · p. 14
• The final equality follows from the assumption hi ∼N(0, δ2).Source: Appendix A · p. 14
• The appendix closes by declaring the proof of Lemma 1 complete.Source: Appendix A · p. 14