Google-EPFL Collaboration
2023 Project Opportunity – Efficient Min Max Formulations With Applications to Derasterization
Research description: Stylus derived inputs (SDI), for example drawings and handwriting can be represented in a vector format (collection of bezier curves) or a rasterized format (collection of pixels on a grid, e.g. from a picture of handwriting). Vector format has the benefit of being losslessly rescalable, while rasterized formats are simpler to create and work with for generalised shapes. As the digitization of our world continues, both representations of SDI become more frequent, either rasterized via e.g. photographs of whiteboards or in vector format, as input from digital pens or drawing software, motivating the development of new methods to de-rasterize (rasterization being a solved problem) and improve vector based input (which can later be used as a guide for generative image models). As part of a parallel collaboration, Google and LIONS@EPFL will develop an ML model to perform derasterization.
2022 Project Opportunity – Equilibrium-based Federated Learning via Online Optimization in Games: Better, Stronger, Faster
Research Description –Federated learning (FL) poses a fundamental challenge to learning algorithms, which is captured by the prevalent networking, communication, and privacy bottlenecks. In fact, research in FL has largely been driven by these constraints bottom up. In stark contrast, our project takes a top-down rethinking of FL, starting from primal-dual reformulations towards the flexible and powerful online learning in games perspective. The two research fields have been evolving largely in isolation despite the strong connections, which this project will bridge through the research questions towards better, stronger, and faster FL.