Post-Doctoral Research Visit F/M Federated Learning under Energy Limit

This post-doctoral position will be supported by the Fed-MALIN project.
Fed-MALIN addresses a number of challenges that arise when Federated Learning (FL) is deployed over the Internet, including privacy, fairness, energy consumption, personalisation, and location/time dependencies.
Fed-MALIN will also contribute to the development of open-source tools for FL and will use them for concrete applications in medicine and crowdsensing.

The position is part of a collaboration between two teams of Inria Lille: Spirals (Self-adaptation for distributed services and large software systems Magnet (Machine Learning in Information Networks).

The context of this project is Federated Learning (FL) where devices have an a priori and known budget for energy consumption. The exact energy consumption of devices is unknown, but can be evaluated by local measurements reported by middleware toolkits, like PowerAPI (http://powerapi.org). The aim is to design and implement online strategies in FL algorithms that are adaptive to the constraints of the energy limit and to the consequences of these constraints. You will study the impact of budgeted limits and energy consumption approximation on the client and the server side. In particular, devices can adapt the amount of information sent to the server and reduce the computational cost of gradients (using, for instance, quantization or sampling data or parameters).
On the server side, it is therefore necessary to mitigate the induced biases due to the unavailability of the devices, the heterogeneity of the collected gradients. These strategies are driven by local information in the first place, but need to be tuned in a collaborative way.

The post-doctoral research activity includes several key steps:

• Study (local) energy consumption measurement. This include the quality and robustness of PowerAPI measurements, the impact of quantization, model sizes, batch sizes, loss function in classical gradient descent-descent algorithms. Model predictions for energy consumption could also be studied and evaluated;
• Study and manage the impact of heterogeneity of gradients at the server level on the convergence and the accuracy in (standard) aggregation steps. Possible strategies to mitigate the induce bias could also depend on auxiliary knowledge communicated by the clients;
• Devise new collaborative approaches for adaptive consumption of the budget across FL iterations.

PhD in computer science, machine learning, or software engineering/distributed computing.
Strong programming skills in Python/Pytorch.
Prior experience in Federated Learning will be an asset.

• Subsidized meals
• Partial reimbursement of public transport costs
• Leave: 7 weeks of annual leave + 10 extra days off due to RTT (statutory reduction in working hours) + possibility of exceptional leave (sick children, moving home, etc.)
• Possibility of teleworking and flexible organization of working hours
• Professional equipment available (videoconferencing, loan of computer equipment, etc.)
• Social, cultural and sports events and activities
• Access to vocational training
• Social security coverage