Lightweight Bayesian Deep Learning Framework for Uncertainty-Aware Decision Making on Edge IoT Devices

The increasing reliance on Internet of Things (IoT) ecosystems has heightened the demand for intelligent decisionmaking at the network edge, where devices often operate under stringent constraints in memory, computation, and energy. Conventional deep learning models deployed on such devices typically emphasize predictive accuracy but fail to incorporate uncertainty estimation, thereby limiting their reliability in safetycritical or dynamic IoT environments. Bayesian deep learning offers a principled mechanism to quantify predictive uncertainty; however, existing methods are computationally intensive and impractical for real-time edge deployment. To address this challenge, this paper proposes a lightweight Bayesian deep learning framework explicitly designed for uncertainty-aware decision making on resource-constrained IoT devices. The framework leverages variational inference and model compression techniques to balance accuracy, uncertainty estimation, and efficiency, ensuring scalability across diverse IoT modalities. Extensive experiments on benchmark IoT datasets and edge hardware prototypes demonstrate that the proposed approach achieves up to 32 % reduction in energy consumption and 27 % faster inference compared to conventional Bayesian methods, while preserving robust uncertainty quantification. The novelty of this study lies in integrating lightweight Bayesian modeling with edge-oriented optimization, bridging the gap between uncertainty-aware AI and the practical limitations of IoT deployment. This contribution enables more reliable, efficient, and adaptive decision-making in real-world edge intelligence scenarios.