Loading...

Artificial Intelligence as a Service in Edge Computing

Pourakbar, Vahid | 2023

12 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 56861 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Shah Mansouri, Hamed
  7. Abstract:
  8. Artificial intelligence (AI) has become an instrumental force in reshaping various sectors, especially in mobile networks. To realize AI as a service (AIaaS) and enable prompt decision-making, edge computing has risen as an innovative response, pushing computation closer to data sources. Although this is promising to reduce the dependency on remote cloud infrastructures and save the mobile devices’ energy, current methodologies display constraints in effectively distributing deep neural network (DNN) inference tasks, leading to possible resource underutilization and performance compromises. To tackle these issues, in this thesis, we study AIaaS at the edge and focus on renowned DNN architectures like AlexNet and VGG16. We model the inference tasks as directed acyclic graphs(DAGs) and introduce an efficient task offloading mechanism. Our goal is to minimize the energy consumption while adhering to latency requirements. To tackle the complexity of this optimization problem, we propose a distinctive approach and utilize a transformer DNN architecture. By training on historical data, we obtain a feasible and near-optimal solution for the problem. Our findings reveal that under conditions with sufficient computing resources, our transformer model matches the performance of established baseline schemes. However, when computing resources at the edge are limited, our model exhibits a 22\% reduction in energy consumption and a significantly decreased task completion fail ratio comparing to existing works
  9. Keywords:
  10. Mobile Edge Computing ; Combinatorial Optimization ; Transformer Network ; Artificial intelligence as a Service (AIaaS) ; Inference Task Offloading

 Digital Object List