Hide menu

IDA Machine Learning Seminars - Spring 2020



Wednesday, February 26, 3.15 pm, 2020

Underwater Machine Learning - Towards Sidescan Sonar SLAM
Nils Bore
, Dept of Robotics, Perception, and Learning, Royal Institute of Technology (KTH)

Abstract: Since the 1960s, sidescan sonar has been an ubiquitous tool in maritime seafloor surveys, and instrumental for example in the discovery of the RMS Titanic in 1985. While limited by the amount of geometrical information contained in its output signal, sidescan still remains essential in small autonomous underwater vehicles due to its affordability and compact size. In the absence of GPS or pre-placed acoustic beacons, sidescan remains one of the few options for positioning these vehicles with respect to the seabed. In our work, we apply recent developments in neural networks to the task of interpreting sidescan sonar images. To build bathymetric maps, networks are trained to reconstruct the geometry of the seabed from incomplete information. To localize, we develop networks to recognize previously seen places. We also learn latent seabed maps that let us reconstruct the signal from a given area. Our vision is to bring these pieces together in order to achieve a long-sought milestone in the AUV community: effective sidescan sonar localization and mapping (SLAM).

About the speaker: Nils Bore received the M.Sc. degree in mathematical engineering from Faculty of Engineering, Lund University, Lund, Sweden, in 2012 and the Ph.D. degree in computer vision and robotics from the Robotics, Perception and Learning lab, KTH Royal Institute of Technology, Stockholm, Sweden, in 2018. He is currently a PostDoc with the Swedish Maritime Robotics (SMaRC) project at KTH. His research interests include robotic sensing and mapping, with a focus on probablistic reasoning and inference. Recently most of his work has been on applications of specialized neural networks to underwater sonar data. In addition, he is interested in system integration for robust and long-term robotic deployments.

Location: Ada Lovelace
Organizer: Per Sidén


Wednesday, April 22, 3.15 pm, 2020

Seeing in to the future. Using self-propelled particle models to aid player decision-making in soccer.
David Sumpter
, Department of Mathematics, Uppsala University.

Abstract: Soccer has some of the most complex team movement patterns of any team sport. Recently, several measurements have been proposed for evaluating the value of dribbles, passes or shots. The next step is to automatically identify the alternative actions available to players both on and off the ball. We address this challenge by building a ‘self-propelled player’ model, simulating attacking roles by maximizing three criteria: pass probability, pitch Impact and pitch control. The model assumes that players can anticipate the movement of the other players on the pitch a few seconds in to the future and maximize the future value of their position. We compared these simulations to player decisions during matches by top-flight men’s teams of Hammarby IF and FC Barcelona. In simulations, we found that the two or three players nearest to the ball tended to optimize the product of pass probability and pitch impact. In a first-team coaching intervention at Hammarby, players re-watched attacking situations in which they had been involved, and were asked to discuss their own actions in comparison with the model. The players often agreed that the model captured complex game patterns, including off-ball actions. The model also recommended runs that the players hadn’t taken, which the players also found realistic and aided discussions. Despite the novelty of these discussions, the players showed a high willingness to engage with them. We further explored how these techniques can be used to provide automated feedback to players within the match cycle. Bio: David Sumpter is professor of applied mathematics and author of Soccermatics (2016), Outnumbered (2018) and The Ten Equations that Rule the World (due 2020). His research, resulting in over 100 publications, covers everything from the inner workings of fish schools and ant colonies, through social psychology and segregation in society, to machine learning and artificial intelligence. He has consulted for leading football clubs and works actively with outreach to schools, industry and the social sector. His talks at Google, TedX, the Oxford Mathematics Public Lecture and The Royal Institution are available online. 

Location: Remotely via Zoom
Organizer: Patrick Lambrix



Page responsible: Fredrik Lindsten
Last updated: 2020-11-27