DeCLIP: Pareto-Efficient Collaborative Inference in Decentralized Physical Infrastructure Networks
IEEE Open Journal of the Computer Society · 2026 · 7 · 1051-1063
DOI: 10.1109/ojcs.2026.3690486 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Open Journal of the Computer Society, 2026. Verfasst von Yao Du und 4 weiteren Autor:innen. Die Arbeit nennt 55 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Blockchain@UBC, Natural Sciences and Engineering Research Council of Canada, Guangdong Pearl River Talent Recruitment Program.
Abstract (Originalsprache)
Large artificial intelligence models (LAMs) face deployment challenges in next-generation networks, which demand ultra-low latency and context-aware intelligence for agentic applications. Traditional cloud-based artificial intelligence (AI) infrastructures centralize distributed data for processing; this incurs prohibitive latency for time-sensitive edge AI tasks. To address the centralized LAM deployment challenge, we propose the use of decentralized physical infrastructure networks (DePINs) to deploy contrastive language-image pre-training (CLIP) models (Radford et al., 2021) in decentralized data environments. Effective inference throughput (EIT) is proposed to optimize Pareto efficiency with diverse accuracy and throughput priorities. We propose decentralized CLIP (DeCLIP) and context chain (CoChain) to maximize EIT for Pareto-efficient CLIP inference. We theoretically derive the CLIP inference error bounds for DeCLIP and CoChain, which respectively achieve empirical accuracy improvements of 15% and 30% on the COCO dataset under data distribution shifts. Notably, the DeCLIP runs 2× faster than the original CLIP on the decentralized data. With time-slotted problem decomposition and reduction, the optimality gap remains below 1% while achieving a 5× speedup over the optimal scheduler.
Förderung
- Blockchain@UBC
- Natural Sciences and Engineering Research Council of Canada
- Guangdong Pearl River Talent Recruitment Program
- Guangdong Pearl River Talent Plan
- UBC PMC-Sierra Professorship in Networking and Communications
Zitieren
APA
Du, Y., Leung, C., Liu, J., Wang, Z., & Leung, V. C. M. (2026). DeCLIP: Pareto-Efficient Collaborative Inference in Decentralized Physical Infrastructure Networks. IEEE Open Journal of the Computer Society, 7, 1051-1063. https://doi.org/10.1109/ojcs.2026.3690486
Vancouver
Du Y, Leung C, Liu J, Wang Z, Leung VCM. DeCLIP: Pareto-Efficient Collaborative Inference in Decentralized Physical Infrastructure Networks. IEEE Open Journal of the Computer Society. 2026;7:1051-1063. doi:10.1109/ojcs.2026.3690486
BibTeX
@article{Du2026,
author = {Du, Yao and Leung, Cyril and Liu, Jing and Wang, Zehua and Leung, Victor C. M.},
title = {DeCLIP: Pareto-Efficient Collaborative Inference in Decentralized Physical Infrastructure Networks},
journal = {IEEE Open Journal of the Computer Society},
year = {2026},
volume = {7},
pages = {1051--1063},
doi = {10.1109/ojcs.2026.3690486}
}
EUVIMED