CEDoc – UM6P – College of Computing: Peer-to-Peer Machine Learning Scheme Full-time

Introduction UM6P

Mohammed VI Polytechnic University is an institution dedicated to research and innovation in Africa and aims to position itself among world-renowned universities in its fields The University is engaged in economic and human development and puts research and innovation at the forefront of African development.

A mechanism that enables it to consolidate Morocco’s frontline position in these fields, in a unique partnership-based approach and boosting skills training relevant to the future of Africa. Located in the municipality of Benguerir, in the very heart of the Green City, Mohammed VI Polytechnic University aspires to leave its mark nationally, continentally, and globally

Context

Machine learning (ML) is a subset of artificial intelligence that focuses on developing algorithms and statistical models that enable computers to learn from and make predictions or decisions based on data without being explicitly programmed. In essence, machine learning algorithms use patterns and insights from data to improve their performance over time, allowing them to identify trends, make predictions, and automate decision-making processes.

In the literature, there are two widely approaches of ML both relying on a central server: centralized and federated learning. In centralized learning (CL), each client sends the data to the server, where the latter uses the data of all the clients to generate a model hoping that it suits all. On the other hand, in the federated learning (FL), instead of sending the data to the central server, each worker carries out a local training phase and sends the generated model to the server. Upon receiving all the models from the workers, the server applies some aggregation techniques to generate a common model that fits all the workers.

Unfortunately, this server poses risks such as being a single point of failure (SPoF) and compromising data privacy and security. Additionally, the one-size-fits-all approach adopted by CL and FL schemes is inadequate as there are real-life use cases where each node has unique characteristics and requirements. Recently, a new learning scheme has been proposed inspired by the peer-to-peer networks, known as the peer-to-peer learning scheme (P2P), where peers having similar characteristics collaboratively train ML models. The advantages of such a learning scheme are that not only it solves the SPoF and data privacy issues, but it has also been shown that such a scheme is extremely energy efficient if trained using efficient devices such as IoT and mobile phones, among others.

Research Objectives

This PhD thesis will be based on P2P learning scheme and address the challenges present both from theoretical and practical perspectives. From theory, such a P2P learning scheme and its accuracy of inference depends largely on the underlying structure of the graph. Hence, new methodologies need to be investigated in dynamically selecting the neighbors of a given peer under the condition that certain criteria are met such as communication bandwidth, availability of reliable data, energy consumption, and many more. Another challenge that this PhD thesis will be tackling is the derivation of energy and resource efficient algorithms for P2P learning scheme.

To this end, tools like roofline can be used to analyze the efficiency of the underlying code and hence will inspire further improvements from software implementation point of view. The third challenge of the corresponding thesis is to implement the devised resource and energy efficient P2P learning algorithms on heterogeneous and resource limited devices such as IoT or mobile phones. In this regard, custom implementations will be needed both for the usage of CPU and GPU in the training. The developed P2P algorithms will be practically implemented and tested within the context of Smart Buildings or Homes by considering the use cases of solar panels and/or Smart Thermostats.

Admission Criteria
• A master’s degree in computer science or equivalent.
• Aptitude for teamwork, problem-solving, and collaborative relationships.
• Strong technical background in at least one of the following areas:

o Model-driven engineering and machine learning in general, and centralized, federated and peer-to-peer learning in specific.

o A sound mathematical background

o Android and Python-based libraries for CPUs and GPUs.

o System modelling and optimization formulation and solving.

References
• Boubouh, Karim, Robert Basmadjian, Omid Ardakanian, Alexandre Maurer, and Rachid Guerraoui. “Efficacy of temporal and spatial abstraction for training accurate machine learning models: A case study in smart thermostats.” Energy and Buildings 296 (2023): 113377.
• Boubouh, Karim, Robert Basmadjian, Omid Ardakanian, Alexandre Maurer, and Rachid Guerraoui. “PePTM: An Efficient and Accurate Personalized P2P Learning Algorithm for Home Thermal Modeling.” Energies 16, no. 18 (2023): 6594.
• Boubouh, Karim, and Robert Basmadjian. “Power Profiler: Monitoring Energy Consumption of ML Algorithms on Android Mobile Devices.” In Companion Proceedings of the 14th ACM International Conference on Future Energy Systems. 2023.
• Boubouh, Karim, Robert Basmadjian, Omid Ardakanian, Alexandre Maurer, and Rachid Guerraoui. “Efficient and Accurate Peer-to-Peer Training of Machine Learning Based Home Thermal Models.” In Proceedings of the 14th ACM International Conference on Future Energy Systems, pp. 524-529. 2023.
• Basmadjian, Robert, Karim Boubouh, Amine Boussetta, Rachid Guerraoui, and Alexandre Maurer. “On the advantages of P2P ML on mobile devices.” In Proceedings of the Thirteenth ACM International Conference on Future Energy Systems, pp. 338-353. 2022.
• Zantedeschi, Valentina, Aurélien Bellet, and Marc Tommasi. “Fully decentralized joint learning of personalized models and collaboration graphs.” In International Conference on Artificial Intelligence and Statistics, pp. 864-874. PMLR, 2020.

UM6P.