Loading...

Accelerating Neural Networks Execution on Resource-constrained Devices

Amiri, Mahdi | 2022

93 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 55646 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Hessabi, Shaahin; Rohban, Mohammad Hossein
  7. Abstract:
  8. The development of deep neural networks is making tremendous progress in various fields, including processing Image, speech processing and other areas. Despite this tremendous achivement, neural networks have a lot of computational overhead and memory access that prevent them from being used in resource-constrained devices. We also know that many neural network applications are of great importance in mobile devices, and it is desirable for us to use their power in this regard. Many efforts have been done at different levels to solve the problem of executing deep neural networks on these devices. In this research, an approach based on offloading is used in which two different small (on the resource-constrained device side) and large (on the cloud side) networks work together to calculate outputs. To this end, a novel training process has been proposed in which any combination of blocks from both networks can work together to achieve the desired result. It has been shown that with this method, we can achieve an accuracy between the accuracy of two fully large and small neural networks (for example, adversarial accuracy of 48.49% to 53.42% and clean accuracy of 80.38% to 84.25% in the CIFAR-10 dataset). Moreover, it is shown that the more blocks of the large network are present in the calculations, the higher the final accuracy of the calculations
  9. Keywords:
  10. Offloading ; Deep Neural Networks ; Knowledge Distillation ; Adversarial Training ; Cloud Environment

 Digital Object List

 Bookmark

No TOC