Abstract

The advancement of deep learning methods has ushered in novel research in the field of computer vision as the success of deep learning methods are irrefutable when it comes to images and video data. However deep learning methods such as convolutional neural networks are computationally heavy and need specialized hardware to give results within a reasonable time. Early-exit neural networks offer a solution to reducing computational complexity by placing exits in traditional networks bypassing the need to compute the output of all convolutional layers. In this thesis, a reinforcement learning-based exit selection algorithm for early-exit neural networks is analyzed. The exit selection algorithm receives information about the previous state of the early-exit network to make decisions during runtime. The state of the early-exit network is determined by the previously achieved accuracy and inference time. A novel feature is proposed to improve the performance of the reinforcement learning network to make better decisions. The feature is based on the input image and attempts to quantify the complexity of every single input. The impacts of adding the new feature and potential performance improvements are documented. The testing is performed on two image classification datasets to record any variance in performance with the dataset. A scenario with the computation of the exit selection algorithm offloaded to a local edge server is also investigated by creating a simple analytical edge computing model. The decision-making rate is varied in this scenario and the potential differences in performance are documented.

Publication Date

11-2021

Document Type

Thesis

Student Type

Graduate

Degree Name

Computer Engineering (MS)

Department, Program, or Center

Computer Engineering (KGCOE)

Advisor

Amlan Ganguly

Advisor/Committee Member

Corey Merkel

Advisor/Committee Member

Sai Manoj

Comments

This thesis has been embargoed. The full-text will be available on or around 12/23/2022.

Campus

RIT – Main Campus

Available for download on Friday, December 23, 2022

Share

COinS