MSMCC: Multi-Small Model Coordination Center for Accelerating Network Intelligence*
Yuhong Huang, Liexiang Yue, Min Zhang, Qingbi Zheng, Na Li, Guangyi Liu, Qixing Wang · 2024
Presently, expansive artificial intelligence models are heralding a revolutionary epoch of technological advancement, with AI-generated content (AIGC), exemplified by ChatGPT, at its vanguard. The relentless evolution of artificial intelligence within networks necessitates exponential growth: on one hand, large model applications impose substantial demands on network bandwidth and real-time interaction; on the other, the extensive training and inference of these models require colossal computational power and prodigious energy consumption to sustain the requisite hardware. This dual exigency not only impinges upon the commercial viability of network large models but also impedes their widespread adoption. To surmount these challenges, this paper proposes a Multi-Small Model Coordination Centre (MSMCC) network architecture that amalgamates large models with networks. The crux of this approach lies in decomposing intricate problems into multiple, more tractable sub-problems, each addressed by distinct sub-models. By synthesizing the outputs of these sub-models to tackle specific categories of problems and furnish definitive solutions, the overall performance and efficiency of the model are augmented. Empirical evaluations have demonstrated that MSMCC can curtail overall inference time by 10-15% compared to the baseline, while maintaining commensurate or superior accuracy.