BISCOTTE

With the enormous amounts of data involved in training a machine learning system, as well as the trend to push artificial intelligence to mobile devices and embedded systems, taking into account computation and memory constraints is of primary interest from the get go when designing machine learning methods. A defining feature of modern approaches to artificial intelligence is that the different type of constraints coming from statistical efficiency (i.e. training data efficiency) and computational efficiency (i.e. memory and compute time efficiency), should be considered simultaneously, not separately. While thinking about efficient algorithms in the two above senses was present from the origins of the field of machine learning, joint consideration of these issues has accelerated significantly in the last decade.The overarching objective of this chair will be to combine expertise and tools from optimisation, statistics, and theoretical computer science to take into account structure hidden in the data, and design provably reliable, statistically efficient, and computationally efficient algorithms exploiting such structure. One key aspect that will be considered is that of adaptivity, i.e. automatic tuning of the statistical models or hyperparameters involved, as well as of the type and amount of computational acceleration (such as parallelization, low dimensional representation), driven by he data itself. For the teaching objectives, a central tenet is to expose masters’ students to relevant parts of the fields of mentioned above (optimisation, statistics, theoretical computer science) and to bring them in a position to understand, control and ultimately shape (and not only to use) the latest developments in the field of artificial intelligence. Interaction between the different fields will be emphasized, so that the student’s thinking and skillset is shaped by a principle of mutual cross-overs, as much as through specialized courses on technical and up-to-date topics.