Harnessing Structure in Optimization for Large-scale Learning – STROLL

The growth and diversification in data collection techniques has led to tremendous changes in the optimization methods used in machine learning. Several research directions have recently proven to be able to scale up to current challenges. Among them, let us focus on two promising trends:
i) Dimension Reduction: identifying pertinent directions in the variable search space and concentrate most computational efforts onto these ones (e.g. variable screening for the lasso problem).
ii) Distributed Computing: computing clusters are now easily available, and learning using local computations on handheld devices has been initiated by Google's federated learning framework.

The success of these topics is partly due to a common denominator: the optimization problem is strongly structured and this structure is used to produce efficient methods. In this project, we are interested in the design and analysis of structure-harnessing optimization methods for large-scale learning