Multi-Source and Multi-Dimensional Information Fusion in Resource-Constrained Edge Networks

Yushi Wang; Zhengxin Yu; Guhan Zheng; Wang Miao; Merouane Debbah

cs.DCOct 9, 2026

Multi-Source and Multi-Dimensional Information Fusion in Resource-Constrained Edge Networks

Authors: Yushi Wang, Zhengxin Yu, Guhan Zheng, Wang Miao, Merouane Debbah

Organizations: School of Computing and Communications, Lancaster University, LA1 4YW, U.K. · School of Communication and Information Engineering, Shanghai University, 200444, China · Department of Computer Science, University of Exeter, EX4 4QF, U.K. · KU 6G Research Center, Khalifa University of Science and Technology, 127788, Abu Dhabi, UAE

Abstract

Multi-source semantic information fusion enables edge networks to support global inference using observations collected by distributed terminals. However, limited communication and computation resources restrict semantic uploading, leaving the edge server with incomplete tokens that differ in modality, semantic granularity, arrival time, and channel condition. The value of these tokens also depends on the complementary or redundant information provided by other terminals. In this paper, we propose a multi-source and multi-dimensional semantic information fusion framework that coordinates terminal uploading and edge-side fusion under resource constraints. We develop Hyper-Slot to combine semantic content with available token attributes through iterative token-to-slot assignment and slot-to-token feedback, producing a compact representation for global inference. To encourage useful semantic uploading, we further design a contribution-based incentive mechanism that uses Hyper-Slot assignments to evaluate information coverage and cross-type complementarity. A bi-level Stackelberg-Nash game determines the server reward pool and terminal transmission-quality strategies while accounting for computation and communication energy costs. Experiments on a tri-modal extension of COCO2014 and image-text retrieval on Flickr30K show that Hyper-Slot improves mAP by 3.66 percentage points and RSum by 15.9 points over the strongest evaluated baselines, respectively, while reducing the number of parameters by more than 80% compared with a naive concatenation-based fusion baseline.

Figures & tables

Fig. 1: Overview of the proposed multi-source and multi-dimensional semantic information fusion framework.

Fig. 2: Structure of the proposed Hyper-Slot module. Semantic tokens are assigned to latent slots using semantic content and token attributes, and iteratively aggregated into a compact representation for global inference.

Method	mAP over random seeds						Efficiency
Method	Seed 42	Seed 123	Seed 456	Seed 789	Seed 1024	Mean $\pm$ Std	Params (M)	GFLOPs	Time (ms)	Memory (MB)
Mean Pool [ 34 ]	0.5431	0.5349	0.6876	0.6964	0.6920	0.6308 $\pm$ 0.0839	1.193	0.0582	0.0300	132.62
Concat MLP [ 33 ]	0.6345	0.6309	0.6386	0.6536	0.6568	0.6429 $\pm$ 0.0116	22.034	0.0999	0.0274	450.40
SCAN [ 17 ]	0.6564	0.6706	0.7599	0.7702	0.7651	0.7244 $\pm$ 0.0560	3.365	0.3364	0.0880	165.53
Slot Attention [ 22 ]	0.7009	0.6969	0.7312	0.7502	0.7407	0.7240 $\pm$ 0.0239	1.860	0.1526	0.0565	143.38
HGNN [ 21 ]	0.6548	0.6469	0.7021	0.7195	0.7108	0.6868 $\pm$ 0.0336	2.183	0.2263	0.0720	147.94

TABLE I: Comparison of mAP and computational efficiency over five random seeds.

Fig. 3: Comparison of per-class mAP across five categories on COCO2014 dataset.

Fig. 4: Visualization of token-to-slot assignments in Hyper-Slot. Colors indicate the assigned slots, and star markers denote the slot anchors.

Fig. 5: Effect of heterogeneity descriptor configurations on mAP.

Fig. 6: Comparison of token selection strategies under different token transmission budgets. The dashed line indicates the full-transmission performance.

Fig. 7: Structural ablation of Hyper-Slot: mAP over training epochs for the compact-only variant, the dependency-only variant, and the full model.

Method	RSum over random seeds						Efficiency
Method	Seed 42	Seed 123	Seed 456	Seed 789	Seed 1024	Mean $\pm$ Std	Params (M)	GFLOPs	Time (ms)	Memory (MB)
VSE++ [ 36 ]	411.2	417.5	412.9	419.0	415.1	415.1 $\pm$ 3.2	2.886	0.1825	0.0330	31.39
Concat MLP [ 34 ]	428.6	433.1	430.5	436.2	432.4	432.2 $\pm$ 2.9	80.484	0.3377	0.0632	330.15
GPO [ 34 ]	446.8	451.5	454.2	456.0	450.9	451.9 $\pm$ 3.5	2.911	0.1826	0.1267	50.52
HGNN [ 21 ]	436.4	443.1	439.8	446.2	441.5	441.4 $\pm$ 3.7	5.018	0.3056	0.1025	49.13
Slot Attention [ 22 ]	458.7	463.4	461.2	467.0	464.8	463.0 $\pm$ 3.2	26.006	1.3289	0.3087	132.91

TABLE II: Comparison of RSum and computational efficiency on Flickr30K over five random seeds.

Fig. 8: Effect of the reward pool on server utility, fusion benefit, total system cost, and equilibrium resource utilization.

Panel of Figure 11 — Fig. 9: Terminal contributions and reward allocation under the uploading strategies

Multimodal Fusion

Multimodal fusion combines information from different data types to improve prediction, perception, and decision-making.

See all 772 papers

SAGE: Training-Free Semantic Evidence Composition for Edge-Cloud Inference under Hard Uplink Budgets
Apr 21, 2026Inhyeok Choi, Hyuncheol ParkEfficient Inference Edge-Cloud Computing
School of Electrical Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, South Korea
FusionSense: Tri-Stage Near-Sensor Learning for Runtime-Adaptive Multimodal Edge Intelligence
May 19, 2026Sanggeon Yun, Ryozo Masukawa, Minhyoung Na +5Efficient Multimodal Inference Edge Computing
University of California, Irvine Irvine, CA, USA · Kookmin University Seoul, South Korea · Shibaura Institute of Technology Saitama, Japan
CST: Collaborative Selective Transmission for Communication-Efficient Multimodal Edge Inference
Aug 22, 2026Hai Chi, Junrui Zhang, Rui Ning +4Communication-Efficient Collaborative Perception Edge Inference
Department of Electrical and Computer Engineering, University of Arizona, Tucson, AZ, USA · Department of Computer Science, Old Dominion University, Norfolk, VA, USA · InterDigital, Conshohocken, PA, USA

Apr 21, 2026J/K move · Enter open · S save

SAGE: Training-Free Semantic Evidence Composition for Edge-Cloud Inference under Hard Uplink Budgets

Inhyeok Choi, Hyuncheol Park

School of Electrical Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, South Korea

Efficient Inference Edge-Cloud Computing Image Classification Edge-Cloud Collaborative Inference

Open paper AskPDF arXiv

Edge-cloud hybrid inference offloads difficult inputs to a powerful remote model, but the uplink channel imposes hard per-request constraints on the number of bits that can be transmitted. We show that selecting transmitted content based solely on attention-based importance, the standard approach in collaborative inference, is inherently limited under hard budgets. Two findings support this claim. First, replacing high-importance units with low-importance but complementary ones improves server accuracy. This shows that what matters is not individual importance but how well the transmitted set covers diverse aspects of the input. Second, spatially uniform selection without any content information achieves competitive accuracy at moderate budgets. This confirms that spatial coverage alone carries independent value. Based on this analysis, we propose SAGE (Semantic Attention-Guided Evidence), a principled, training-free method that combines importance filtering with embedding-diversity sampling. SAGE achieves 93% of the server ceiling in offloaded accuracy while transmitting fewer than half of the available evidence units on ImageNet-1K, substantially outperforming importance-only composition.

Abstract

Multimodal Fusion

Explore similar work

SAGE: Training-Free Semantic Evidence Composition for Edge-Cloud Inference under Hard Uplink Budgets

FusionSense: Tri-Stage Near-Sensor Learning for Runtime-Adaptive Multimodal Edge Intelligence

CST: Collaborative Selective Transmission for Communication-Efficient Multimodal Edge Inference