Source-linked AI summary
Multi-view Semantic-aware Resource Allocation for Edge AI
Jiawei Luo, Hussein Saleh, Mohamad Assaad
TL;DR
Wireless resource scarcity makes task-relevant communication important, while existing OIB handles only a single view. The paper extends OIB to MGIB and formulates joint compression, mode, and PRB allocation for multi-user Edge AI; its LR scheme achieves near-optimal utility and twice the PRB efficiency of Average+Random.
Problem
Scarce spectrum motivates task-relevant communication, but OIB supports only a single view and existing general variational solutions are computationally heavy.
Source: §I · p. 1Method
The paper extends OIB to MGIB for complementary-view fusion and formulates joint compression-level, transmission-mode, and PRB assignment as a MILP solved with Lagrangian Relaxation.
Source: §I · p. 1Results
The LR scheme achieves near-optimal utility across the PRB sweep and reaches ≈0.95 utility at R = 35 PRBs, while Average+Random requires R = 60 for the same utility, a 2× PRB efficiency advantage.
Source: §V · p. 5, §VI · p. 5Takeaways & Limitations
The framework combines multi-view semantic compression with resource allocation for multi-user Edge AI while preserving closed-form GIB tractability.
Source: §I · p. 1, §VI · p. 5
Abstract
from arXiv · showhide
Semantic communication is emerging as a pivotal paradigm to overcome wireless resource scarcity in 6G networks. This paper proposes a novel end-to-end framework for semantics-aware resource allocation in multi-user Edge AI. We extend the Opportunistic Information Bottleneck (OIB) to a multi-view setting, termed the Multi-view Gaussian Information Bottleneck (MGIB), by fusing complementary source views without variational approximation or network retraining. A resource allocation framework is then formulated for joint compression-level selection, transmission mode selection (i.e., the number of active sources per group), and PRB assignment. The problem is formulated as a Mixed-Integer Linear Program (MILP), solved efficiently via a proposed Lagrangian Relaxation (LR) algorithm, and benchmarked against the NP-hard optimal solution. Simulations demonstrate that the proposed LR scheme achieves near-optimal utility, and doubles PRB efficiency versus average allocation.
I. INTRODUCTION
The introduction motivates task-relevant communication under scarce spectrum and identifies limits in existing compression and multi-view methods. It proposes MGIB and a resource-allocation framework for multi-user Edge AI; the IB background frames compression as preserving information relevant to a target task.
I. INTRODUCTION
- Finite spectrum and energy motivate goal-oriented communication that transmits information meaningful to the task rather than bits without regard to relevance,,.The paper places this shift in the context of 6G intelligent applications.Source: §I · p. 1
- OIB enables tractable compression without retraining but handles only one view, while VIB is computationally heavy and multi-view alternatives lack resource-linked rate control or rely on variational bounds,,,,,.Classical GIB is closed-form but restricted to Gaussian variables.Source: §I · p. 1
- MGIB extends OIB by analytically constructing view-specific compression matrices from Gaussian layers in a pre-trained network, without variational approximation or network retraining,.The paper presents this as a tractable way to fuse complementary views.Source: §I · p. 1
- The proposed system jointly selects compression level, the number of active sources per group, and PRB assignments in a MILP [7].The paper proposes an LR algorithm that decomposes the problem into per-group subproblems and reports near-optimal utility across PRB regimes.Source: §I · p. 1
A. Information Bottleneck (IB)
- The IB principle compresses a source while retaining information relevant to a target task, trading off compression and task relevance through a Lagrangian.General solutions may require intractable mutual-information estimation; VIB approximates the terms with variational bounds at computational cost.Source: A. Information Bottleneck (IB) · pp. 1–2
B. Gaussian Information Bottleneck (GIB)
For jointly Gaussian source and target variables, GIB admits a closed-form linear compression mapping. Its eigenvalue thresholds provide a tunable way to vary the number of extracted features.
B. Gaussian Information Bottleneck (GIB)
- When source and target variables are jointly Gaussian, the IB problem has a closed-form solution with an optimal linear compression mapping.The mapping includes an independent additive Gaussian noise vector.Source: B. Gaussian Information Bottleneck (GIB) · p. 2
- The passage reports eigenvalues {λi} and eigenvectors {vi} from the stated step, with the eigenvalues sorted in ascending order.Source: B. Gaussian Information Bottleneck (GIB) · p. 2
- As β passes each critical threshold βc_i, a row is added to the compression matrix, increasing the extracted feature dimension n_z without variational approximation or network retraining [5].The threshold determines when feature i is activated.Source: B. Gaussian Information Bottleneck (GIB) · p. 2
C. Opportunistic Information Bottleneck (OIB)
OIB applies closed-form GIB compression at an approximately Gaussian layer in a pre-trained network, but is limited to a single view. The paper extends this approach to multi-view fusion and motivates it as a compact use of complementary sources under PRB scarcity.
C. Opportunistic Information Bottleneck (OIB)
- OIB identifies an approximately Gaussian intermediate layer in a pre-trained network and applies closed-form GIB to create a tunable feature extractor without retraining [5],.The layer is typically identified after a DFT transformation.Source: C. Opportunistic Information… · p. 2
III. SYSTEM MODEL AND THE OPTIMIZATION PROBLEM
- Because PRBs are scarce relative to the groups, transmitting raw data from all sources is infeasible; multi-view fusion instead combines complementary views into a richer, more compact representation.The system allocates R PRBs across K groups, each containing C sources.Source: §III · p. 2
A. MGIB Framework
- MGIB extends OIB by applying the GIB solution independently to each view’s approximately Gaussian layer, yielding view-specific compression and reconstruction matrices for feature dimensions governed by β.The layers are identified within a shared pre-trained network after a DFT transformation.Source: A. MGIB Framework · p. 2
1) Offline Stage:
The framework prepares view-specific feature extractors offline, then compresses and transmits features for reconstruction, attention-based fusion, and classification online.
1) Offline Stage:
- Offline, the pipeline prepares data and applies DFT, trains a reference model, generates view-specific regression targets, solves a GIB generalized eigenvalue problem per view, and stores compression matrices for desired feature dimensions.An attention module is trained on reconstructed features while the classifier is frozen.Source: 1) Offline Stage · p. 2
2) Online Stage:
- Online, sources linearly compress DFT-transformed inputs for transmission; the edge server reconstructs the features with L-MMSE, fuses them using learned attention, and passes them to the classifier.The attention module generates an element-wise map to weight the views based on their quality and complementarity.Source: 2) Online Stage · p. 2
B. Resource Optimization Problem Formulation
The formulation jointly chooses compression level, transmission mode, and PRB assignment for each group. It represents these choices with binary variables and constrains transmission rates by available PRB capacity.
B. Resource Optimization Problem Formulation
- Each group jointly selects a compression level, a transmission mode determining active sources, and PRBs assigned to those sources.The compression level sets the required transmission rate, while mode and channel quality affect PRB needs and rate feasibility.Source: B. Resource Optimization Problem… · pp. 2–3
- Binary variables encode the compression-mode pair, source activation, and assignment of PRBs to sources.The networkwide objective maximizes total utility.Source: B. Resource Optimization Problem… · p. 3
- Constraints enforce one compression-mode pair per group, match active-source counts to modes, and restrict PRB assignments to active sources.PRBs are exclusive across sources, and assigned capacity must support each active source’s compressed transmission.Source: B. Resource Optimization Problem… · p. 3
IV. SEMANTICS-AWARE RESOURCE ALLOCATION - PROPOSED SOLUTION
The proposed Lagrangian Relaxation exploits the MILP’s separable structure: relaxing the only constraint coupling all groups yields independent per-group subproblems solved by a greedy procedure.
IV. SEMANTICS-AWARE RESOURCE ALLOCATION - PROPOSED SOLUTION
- Relaxing C4, the only constraint coupling all K groups, decomposes the problem into K independent subproblems solved by a greedy procedure.The approach is proposed to solve the resource allocation problem efficiently.Source: §IV · p. 3
1) Relaxation:
The relaxation prices PRBs so each group can optimize independently; a subgradient dual update adjusts prices according to PRB overdemand.
1) Relaxation:
- Relaxing C4 assigns each PRB a nonnegative dual price, and each group subproblem maximizes semantic utility minus the price of assigned PRBs.The subproblems retain C1, C2, C3, and C5 for that group.Source: 1) Relaxation · pp. 3–4
- For each candidate compression-mode pair, the channel-aware greedy procedure assigns PRBs by bits-per-price ratio until C5 is satisfied, then selects the pair with highest net utility.Source: 1) Relaxation · p. 4
2) Dual Update:
- The subgradient measures PRB overdemand, and prices are updated with an aggressive diminishing step for the first 50 iterations followed by the Polyak formula.The step size uses the best feasible utility found so far and ρ ∈ (0, 2).Source: 2) Dual Update · p. 4
3) Rounding:
A two-pass rounding procedure repairs conflicts in the final solution and adjusts compression or disables transmission when feasibility cannot be recovered.
3) Rounding:
- Rounding awards each contested PRB to the group with highest utility, then lets groups repair C5 from unused PRBs sorted by channel quality.If available bits are insufficient, compression is reduced to the highest supportable level; groups that remain infeasible use j = 0.Source: 3) Rounding · p. 4
V. EXPERIMENTAL RESULTS AND ANALYSIS
The experiments validate MGIB on MNIST Rotation+Noisy and show that the proposed LR scheme achieves near-optimal utility while using PRBs more efficiently than Average+Random.
Evaluation setup
- The resource-allocation evaluation uses K = 15 groups with C = 4 sources per group and sweeps R from 5 to 70 PRBs over 500 independent Rayleigh-fading runs.Results are reported as mean ± standard deviation.Source: §V · p. 4
MGIB validation
- MGIB’s multi-view fusion clearly outperforms single-view baselines on MNIST Rotation+Noisy, and Table II reports state-of-the-art accuracy against established methods.The experiments use rotated and noisy MNIST views; Table II is labeled as an SOTA comparison on that dataset.Source: §V · p. 4, Figure 2 · p. 5, Table II · p. 5
Resource allocation results
- At R = 35 PRBs, LR reaches ≈0.95 utility, while Average+Random requires R = 60 to reach the same level—a 2× PRB efficiency advantage.The utility gap is +0.08 at R = 30 and widens to +0.12 at R = 35 before narrowing as both schemes saturate beyond R = 50.Source: §V · p. 5
- Across the full PRB sweep, LR closely tracks the MILP optimal benchmark and achieves near-optimal utility.Fig. 3 compares LR with the MILP optimum and Average+Random.Source: §V · pp. 4–5
VI. CONCLUSION
The paper presents a semantics-aware resource-allocation framework for multi-user Edge AI, combining multi-view OIB with joint compression, mode selection, and PRB assignment. Its LR algorithm achieves near-optimal utility and a 2× PRB efficiency advantage over Average+Random.
Conclusion
- The framework extends OIB to multiple views while preserving GIB’s closed-form tractability.It formulates joint compression, mode selection, and PRB assignment as a MILP.Source: §VI · p. 5
- The LR algorithm decomposes the problem into per-group subproblems and achieves near-optimal utility with a 2× PRB efficiency advantage over Average+Random.Fig. 3 compares utility for LR, the MILP optimum, and Average+Random.Source: §VI · p. 5, Figure 3 · p. 5