Towards Integrated Fine-tuning and Inference when Generative AI meets Edge Intelligence

TL;DR

The paper addresses the mismatch between resource-intensive Generative AI models and resource-constrained edge environments. It introduces GaisNet, a collaborative cloud-edge-end framework that enables data-free bidirectional knowledge relay to support integrated fine-tuning at the edge and inference across distributed devices, using HFSL-based edge fine-tuning and SL-based inference with parameter-efficient techniques. The authors formalize the architecture, discuss major design challenges (model splitting, clustering, service trade-offs, and user groups), and validate the approach with experiments on transformer-based vision models, demonstrating gains from pre-training, fine-tuning, and efficient training/inference, as well as the impact of non-IID data and clustering. They also outline privacy, theoretical, and incentive challenges, offering directions for deploying GAI-EI integration in 6G-edge ecosystems with practical implications for latency, privacy, and energy efficiency.

Abstract

The high-performance generative artificial intelligence (GAI) represents the latest evolution of computational intelligence, while the blessing of future 6G networks also makes edge intelligence (EI) full of development potential. The inevitable encounter between GAI and EI can unleash new opportunities, where GAI's pre-training based on massive computing resources and large-scale unlabeled corpora can provide strong foundational knowledge for EI, while EI can harness fragmented computing resources to aggregate personalized knowledge for GAI. However, the natural contradictory features pose significant challenges to direct knowledge sharing. To address this, in this paper, we propose the GAI-oriented synthetical network (GaisNet), a collaborative cloud-edge-end intelligence framework that buffers contradiction leveraging data-free knowledge relay, where the bidirectional knowledge flow enables GAI's virtuous-cycle model fine-tuning and task inference, achieving mutualism between GAI and EI with seamless fusion and collaborative evolution. Experimental results demonstrate the effectiveness of the proposed mechanisms. Finally, we discuss the future challenges and directions in the interplay between GAI and EI.

Figures (8)

• Figure 1: Transformer-based foundation model and parameter-efficient fine-tuning.
• Figure 2: Parameter-efficient inference from the communication perspective.
• Figure 3: The architecture of the proposed GaisNet.
• Figure 4: The Framework of HFSL-based Model Fine-tuning.
• Figure 5: The Framework of SL-based Task Inference.