Meta-Learning Driven Hyperparameter Optimization for Deep Neural Networks in Resource-Constrained IoT Environments

The high rate of the Internet of Things (IoT) development has prompted the implementation of deep neural networks (DNNs) on edge devices that have stringent memory, computation, and power restrictions. Traditional methods of hyperparameter optimization (HPO) like grid search and Bayesian optimisation are computationally expensive and cannot be used in real-time in resource limited environments. This paper suggests a meta-learning based hyperparameter optimization framework that is specifically optimized to optimize the use of IoT-based DNNs. Using the adaptive search strategies and leveraging previous knowledge, the framework minimizes training, reduces energy usage and keeps a high level of predictive accuracy. Explicit assessments of benchmark IoT information and edge prototypes indicate that the suggested approach converges faster (up to 35 % lesser than traditional methods of HPO) and uses less energy (reduced by 28 % than conventional methods of HPO) and preserves comparable or even high quality. This novelty of this work is its combination with meta-learning and energyconscious optimizing of hyperparameters, which allows scalable, efficient, and sustainable intelligent systems to be created that apply to the Internet of Things.