A Lightweight Cloud-Edge Collaborative Intelligence Inference Framework With Runtime Dynamic Optimization for Resource-Constrained Consumer Electronics

The proliferation of Internet of Things (IoT) and embedded computing has led to widespread deployment of smart consumer electronics requiring edge-based Artificial Intelligence (AI) capabilities. However, the heterogeneous nature of sensing data and dynamic edge environments poses significant challenges for efficient model inference on resource-constrained devices. To address these challenges, this paper presents a lightweight collaborative inference framework designed for consumer electronics. First, we formulate the inference optimization problem as a Mixed-Integer Nonlinear Programming (MINLP) problem, considering channel pruning, early exit and cloud offloading decisions to optimize the trade-off between accuracy and computational cost. Second, we propose a selective model activation mechanism based on Markov Decision Process (MDP), which employs a recursive self-attention mechanism to dynamically track inference budgets and guide decision-making through encoder-decoder architectures. The mechanism integrates entropy regularization during training to ensure robust and diverse execution paths. Comprehensive experiments demonstrate that our framework achieves 65.50% reduction in model parameters and 80.68% reduction in inference Floating Point Operations (FLOPs) while maintaining accuracy loss within 0.81% of the original model, making it suitable for real-time AI applications on resource-constrained consumer electronics.