Sorin Liviu JURJ, Public Dissertation of PhD Thesis

  Date and Time
Friday, October 30, 2020 - 12:00
  Location
online

Thesis Title: " Powering and evaluating deep learning-based systems using green energy "

Author: Sorin Liviu JURJ

PhD Board:

  • Chair: Professor Habil. Dr. Eng. Marius MARCU (Politehnica University Timisoara)
  • PhD Supervisor: Professor Dr. Eng. Mircea VLĂDUŢIU  (Politehnica University Timisoara)
  • Scientific Referees:
    • Acad. Professor Dr. Eng.  Mircea PETRESCU (University Politehnica of Bucharest)
    • Professor Dr. Eng. Liviu Cristian MICLEA (Technical University of Cluj-Napoca )
    • Professor Dr. Eng. Nicolae ROBU (Politehnica University Timisoara)

Thesis Summary:

In recent years, advancements in the field of Artificial Intelligence (AI), especially regarding Deep Learning (DL) algorithms, grew at a rapid pace and will continue this trend for the years to come.

However, due to the fact that these algorithms require a huge amount of time, energy, data, and processing power, their impact on the environment is a defining issue. To solve this problem, considering recent „Green AI” efforts that focus on the energy efficiency of AI systems, we propose four novel environmentally-friendly metrics for evaluating the performance of DL models and systems based not only on their accuracy but also on their energy consumption and cost.

In this Ph.D. thesis, we developed and implemented methods for solving the abovementioned problems by first implementing different novel DL applications that solve different problems related to fraud and security. Then, because we observed that a real-time DL-based system consumes more energy than their non-real-time counterparts, we decided to not only run the same implementation on a platform that consumes 5x less energy, but we also wanted to not pay for this energy consumption. We achieved this by considering the use of green energy and by constructing a novel dual-axis solar tracker that is based on the Cast-Shadow principle and which was later modified with minimal costs. We demonstrated in that our solar tracker is efficient and, to the best of our knowledge, for the first time in literature, that it is possible to completely use solar energy for powering a real-time DL-based system when running inference.

The structure of the doctoral thesis comprises an introductory chapter, a chapter with theoretical context, five chapters dedicated to the presentation of the research carried out and the results obtained, a final chapter dedicated to conclusions, personal contributions and future directions as well as a bibliographic list (with 225 titles consulted and cited). The doctoral thesis spans 192 pages, with the research being supported graphically and synthetically by 91 figures and 39 tables.