AI4Research Seminar: Co-design in fundamental science, end-to-end optimisation of experiments with deep learning

2012 was not only the year of the Higgs boson discovery, powered by machine learning techniques; it was also the year when deep learning showed its power in achieving super-human performance in image recognition tasks. Since then, fundamental science has been using deep learning for analysis tasks across the board. However, this leaves uncovered a crucial ingredient in the success of an experiment: its design. The complexity of particle detectors has in the past discouraged attempts to produce a full optimisation of their design, which can only be effective if it integrates a full model of the inference extraction from the produced data. However, optimised co-design of hardware and software is precisely the second step in a full use of deep learning in fundamental science.

In this seminar Tommaso will describe the present status of this new area of research at the crossroads of physics and computer science and its future prospects, and give one example from an astroparticle physics use case.