Loading...

AI as a Service via DNNs Coordination on Edge

Malekim, Alireza | 2024

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 57454 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Hosein Khalaj, Babak; Shah Mansouri, Hamed
  7. Abstract:
  8. As artificial intelligence (AI) applications continue to proliferate, the demand for deep neural network (DNN) models is on the rise. Many applications employing DNNs are sensitive to processing time or require significant processing power for their inputs. One current method for providing DNN-based services is through edge computing ser- vices. These services have less processing power compared to cloud services but can offer greater accessibility. While DNN models deployed at the edge offer promising prospects for delivering AI services with minimal latency, their potential for collabora- tion remains largely unexplored. In this thesis, we propose that DNN service providers pool their computing resources and model parameters, enabling other DNNs to offload computations without replication. We introduce a novel algorithm named Coordinated DNNs on Edge (CoDE), designed to facilitate cooperation among DNN services by establishing novel inference pathways. CoDE aims to identify the optimal pathway, maximizing the potential reward, which encompasses inference speed and model ac- curacy. With CoDE, DNN models can create new inference paths utilizing their own parameters or those of other models. Additionally, an algorithm is proposed to estimate the optimal path with less search time by predicting the inference accuracy of paths. We assess CoDE’s performance through numerical experiments, demonstrating a 40% increase in inference speed while maintaining a marginal average accuracy reduc- tion of only 2.3%. Experimental results indicate that CoDE achieves superior precision using a simpler model compared to state-of-the-art methods and enhances inference speed
  9. Keywords:
  10. Artificial intelligence as a Service (AIaaS) ; Edge Computing ; Service Quality ; Deep Neural Networks ; Service Coordination ; Reliability ; Path Finding

 Digital Object List