Halima Bouzidi's thesis defense

Summary

Neural networks (NNs) have become a dominant force in the world of technology. Inspired by the human brain, their complex design enables them to learn patterns, make decisions and even predict future scenarios with impressive accuracy. RNs are widely deployed in Internet of Things (IoT) systems, further enhancing the capabilities of interconnected devices by giving them the ability to learn and self-adapt in a real-time context. However, the proliferation of data produced by IoT sensors makes it difficult to send them to a Cloud center for processing. Consequently, processing data closer to its origin, in Edge, enables real-time decisions to be made, thus reducing network congestion.

The integration of RNs with the Edge in IoT systems enables more efficient and responsive solutions, ushering in a new era of Edge AI. Nevertheless, deploying RNs on resourced hardware platforms presents a multitude of challenges. (i) The inherent complexity of RN architectures, which require significant computing and memory capacities. (ii) The limited energy budget characterizing Edge hardware devices, which makes it impossible to support complex RNs, drastically reducing system uptime. (iii) The challenge of ensuring harmony between the design of RNs and hardware devices. (iv) The lack of adaptability to the dynamic execution environment and the complexities of the data to be processed.

To alleviate these problems, this thesis aims to establish innovative methods that extend traditional NR design frameworks (NAS for Neural Architecture Search) by integrating the contextual characteristics of the hardware and runtime environment. Our methods help contribute to the realization of an end-to-end design framework for RNs on Edge hardware devices. They thus make it possible to take advantage of several optimization paths at software and hardware level, improving the performance and efficiency of Edge AI.

Keywords

Hardware-aware Neural Architecture Search, Dynamic Inference, DVFS, Edge AI, Performance Prediction, HW/SW Co-optimization.

Jury composition

Prof. Tulika Mitra, Professor, National University of Singapore, NUS - Rapporteur
. Prof. Olivier Sentieys, Professeur, IRISA, Université de Rennes - Rapporteur
Dr. Nicolas Ventroux, Thales Research & Technology (TRT) - Examiner
Prof. Clarisse Dhaenens, Professor, Thales Research & Technology (TRT) Prof. Clarisse Dhaenens, Directrice du laboratoire CRiSTAL, Université de Lille - Examiner
Prof. Smail Niar, LAMIH, Professor, UPHF/INSA, Thesis Supervisor
Prof. El-Ghazali Tali, Professor, UPHF/INSA - Examiner Prof. El-Ghazali Talbi, Professeur, Laboratoire CRiSTAL, Université de Lille, Co-Encadrant
Dr. Hamza Ouarnoughi, LAMIH, Professeur, UPHF/INSA, Thesis Director Dr. Hamza Ouarnoughi, Associate Professor, LAMIH, UPHF/INSA, Co-Encadrant
Prof. Abdessamad Aitou, Professor, LAMIH, UPHF/INSA, Co-Encadrant Prof. Abdessamad Ait El Cadi, Professor, LAMIH, UPHF/INSA, Guest