KADAKIA-THESIS-2021.pdf (2.68 MB)
Download fileENOS: Energy Aware Network Operator Search
thesis
posted on 2021-05-01, 00:00 authored by Yuti Devang KadakiaUntil recent times, DNNs have been designed using conventional high-precision operator which gives maximum accuracy but is not energy efficient. With the invent of new computationally efficient correlation operators, accuracy deterioration becomes a problem. In addition, all DNNs use a single correlation operator for end-to-end processing. In these cases, using only a high precision operator might not be energy efficient and using alternate operators might affect the accuracy. Thus, an optimal solution would be to systematically combine these operators to maintain the energy-accuracy tradeoff. This is the main goal of my framework called ENOS : Energy Aware Network Operator Search where if the energy constraints for an application is established, the operator sequence for different layers can be automatically synthesized while maximizing the prediction accuracy.
History
Advisor
Trivedi, Amit RanjanChair
Trivedi, Amit RanjanDepartment
Electrical and Computer EngineeringDegree Grantor
University of Illinois at ChicagoDegree Level
- Masters
Degree name
MS, Master of ScienceCommittee Member
Cetin, Ahmet Enis Rao, WenjingSubmitted date
May 2021Thesis type
application/pdfLanguage
- en