Tuesday, 15th of January14h30 (amphi Shannon, 660 building) (see location)
Title: The TrackML challenge: concept, methods and approaches
AbstractThe TrackML challenge aims at recovering the trajectory of the (10 K) particles that create (100 K) 3D points in a high-energy physics tracker detector. A dataset with an extremely large number of events has been created for the challenge using an accurate simulator, thus giving access to the ground truth of particle trajectories. In this seminar we address this problem from a Machine Learning point of view. After describing the problem in abstract terms, we see how it can be viewed from various vantage points (either clustering, or image processing techniques like Hough transforms, or combinatorial approaches), and describe the advantages and limitations of the classical methods that address these. We finally have a look at the results from the accuracy phase on the challenge, along with the concrete approaches that gave the best results. We comment on their relevance both for the physicist, and ML practioners communities.
Contact: guillaume.charpiat at inria.fr
All TAU seminars: here