Deep learning, a subset of AI, has become a prevalent technique in many domains, including computer vision, natural language processing, and robotics. However, implementing deep learning algorithms requires significant computational resources, resulting in high energy consumption. This is particularly crucial for resource-constrained systems, such as IoT Embedded Systems (ESs), where energy efficiency is of paramount importance. To address this, we present the use of computational self-awareness to dynamically adapt machine learning algorithms at runtime to minimize energy consumption. Our suggested approaches focus on utilizing approximation techniques as a key mechanism for adaptivity. This can lead to substantial energy savings of up to 2.5 times compared to traditional non-adaptive algorithms. Additionally, our presented approaches enable efficient use of resources by intelligently adapting algorithms at runtime, enabling systems to operate at optimal performance while maintaining energy efficiency. Our proposed methods offer a viable solution for addressing the energy consumption challenge in deep learning, particularly in IoT and embedded systems.
Check it in Zenodo: https://zenodo.org/record/7331509
