Artificial intelligence and machine learning applications on edge devices for digital carbon footprint optimization

The advent of artificial intelligence (AI) applications has profoundly impacted numerous facets of human existence, precipitating substantial transformations and innovations across diverse sectors, particularly in the context of rapid technological advancements. The integration of deep learning (DL) models with Internet of Things (IoT) devices has engendered widespread and profound effects, manifesting in numerous domains such as smart homes, innovative smart cities, sophisticated healthcare systems, advanced transportation networks, and revolutionary industrial automation processes. These remarkable advancements have enabled substantial progress across these varied domains, offering significant advantages in terms of enhanced energy efficiency, seamless automation, and a greatly improved user experience that elegantly adapts to the specific needs and preferences of individuals. This dynamic and synergistic relationship between DL and IoT is fundamentally revolutionizing how we interact with technology in our daily lives, making our environments smarter and more responsive. However, these notable technological advances have also unfortunately been accompanied by a variety of significant environmental impacts that cannot be overlooked. Specifically, the substantial energy consumption associated with deep learning models that are being executed on Internet of Things (IoT) devices and the resulting carbon footprint produced have become a major concern in recent years, drawing the attention of researchers and policymakers alike. This pressing issue has prompted numerous research studies aimed at understanding and mitigating these adverse effects, as addressing such challenges is essential for the sustainability and longevity of these advancements in technology. The primary objective of this thesis is to conduct a comprehensive comparative study and evaluation focused on the energy consumption, performance metrics (specifically latency and accuracy), and the overall carbon footprint associated with popular deep learning models that are widely used within the industry. These models include MobileNetV2, ShuffleNetV2, SqueezeNet, and ResNet18. The evaluation will be meticulously performed on four distinct IoT platforms, each representing different hardware architectures that vary in capability, such as the Raspberry Pi 5 (which utilizes a CPU), the NVIDIA Jetson Xavier NX (designed with a powerful GPU), the Google Coral USB Accelerator (featuring an efficient TPU), and the Khadas VIM3 Pro (equipped with a specialized NPU). Furthermore, this research will delve into a thorough understanding of the impact of various optimization techniques that can significantly enhance the performance and sustainability of these models. These techniques will include quantization, pruning, and knowledge distillation, with a specific focus on how they systematically affect the energy consumption and the carbon footprint associated with the aforementioned models. This rigorous assessment will be conducted within the practical context of a real-world smart home application, particularly emphasizing sophisticated energy consumption monitoring systems designed to improve sustainability. By integrating these techniques and evaluating their effects in detail, this study aims to provide valuable insights into the optimization of deep learning models for sustainability and efficiency within increasingly prevalent IoT environments. The primary goal of this comprehensive study is not only to thoroughly benchmark but also to critically evaluate the energy consumption of widely utilized deep learning models, such as SqueezeNet and ResNet18, alongside their associated latency and accuracy as essential performance metrics. Moreover, this extensive study will also investigate how these prominent deep learning models are influenced by variations in energy consumption and their corresponding carbon footprint, which is becoming an ever-important concern in today's environmentally conscious world. A critical and significant facet of this research further entails the evaluation and comparison of a variety of powerful optimization techniques, including but not limited to quantization, pruning, and knowledge distillation. This will all be done in the context of a practical, real-world example: a smart home application that is specifically focused on monitoring energy consumption and optimizing usage patterns to foster sustainability and efficient resource management. The optimization process harnesses a variety of effective techniques which include quantization, pruning, knowledge distillation, and weight clustering. The primary implementation of these strategies is specifically aimed at achieving a markedly greater level of energy efficiency. Furthermore, a thorough and comprehensive analysis has been meticulously conducted to explore the wide-ranging impact of these optimization techniques on energy consumption, processing time, as well as model accuracy. This in-depth analysis has been undertaken to identify the most optimal balance between energy efficiency and high performance in Internet of Things (IoT) devices, aiming to elucidate the interplay between computational demands and environmental responsibilities. The data obtained through these advanced methods is anticipated to provide significant and valuable insights, which will be beneficial for both academic researchers and industrial applications alike, fostering a collaborative effort toward a sustainable technological future. This study presents a comparative analysis of the effects of optimization techniques applied to hardware components such as NPU, TPU, and GPU. A comprehensive discussion is provided on the most effective solutions in terms of environmental impacts. The study's findings are conclusive in demonstrating that artificial intelligence models can effectively reduce the carbon footprint without compromising performance and energy efficiency. Consequently, this study underscores the significance of reducing the environmental impact of IoT devices, thereby paving the way towards a sustainable digital future. Achieving this balance between technological advancement and environmentally friendly approaches is imperative.