Running the Numbers at Sloan 2020
Members of the Sportlogiq team will be attending the 2020 edition of MIT’s Sloan Sports Analytics Conference in Boston from March 6th to 7th. The conference provides a forum for industry professionals to discuss the increasing role of analytics in the sports industry.
This year, our very own Michael Horton will present his research paper and poster on “Learning Feature Representations from Football Tracking” on Thursday, March 5th at 12:45pm. Michael is a machine learning researcher on Sportlogiq’s innovation team, SLiQ Labs.
Earlier this year at AFCA, Sportlogiq announced a partnership with Telemetry Sports to bring NFL level analysis to NCAA football programs. As part of the agreement, Sportlogiq is providing unique data and contextual insights to power Telemetry’s premier NCAA Football product, which is changing the landscape in college football. Michael’s presentation highlights just some of the groundbreaking work we’re doing in the sport.
Ahead of the conference, we had the chance to sit down with Michael and talk to him about his path to Sportlogiq, how his work in machine learning can be tied to sports, and the conference.
“About 8 years ago, I was working in finance when I made the decision to go back to university and get a PhD. I studied computational geometry, specifically the analysis of piecewise-linear curves in space, and it turns out that the trajectories captured by player tracking systems are examples of such curves, and is thus something that can be applied in the work I’m doing at the company now.”
“After writing a few papers on that, I caught the attention of Sportlogiq’s CTO, Mehrsan Javan. He continued to keep a close eye on my work and ultimately offered me a position here at Sportlogiq.”
“I believe it has already changed the landscape significantly. Computer vision systems are able to capture tracking data and the locations of all the players during games. That’s driven a huge part of the analysis in the past 10 years in football, soccer, basketball…”
“Tracking data is a great data resource but it needs to be analyzed and we need to create game models to describe what is happening during the play. Machine Learning and in particular deep neural networks have been shown to be effective tools for this.”
“Looking forward, 3D pose estimation, where the location of each joint of the player is captured, is going to open up a whole new set of possibilities for analysis. We can analyze a player’s orientation, the direction they are facing, where they’re shifting their body weight, and other details that will help give us a much richer understanding of the game.”
“Just from being there last year, there’s obviously plenty of interesting presenters from various backgrounds that are thought leaders in the industry. Of course, I am looking forward to the research paper presentations and hearing about the various research directions that folks are taking.”
“One of the best talks I attended last year was from Mike Leach, Mississippi State’s head coach. Just having an expert in his field describing the level of detail in his day to day work and analysis was extremely impressive. The work that goes into something like video analysis, where they’re looking at specific details such as the direction of a player’s knee really struck me. Having 3D pose estimation will enable tools and information to help coaches leverage these sorts of insights.”
“The paper I’m presenting is looking at football, and applying machine learning directly to raw football tracking data. These are two major innovations in sports analytics in the last few years. 1. Provision of tracking data and 2. Use of machine learning and deep learning to make sense of that data.”
“But there’s been a mismatch in using tracking data as an input to machine learning algorithms that are quite particular about the structure of the data they can input. What this work does is it builds a model that deals with that mismatch in a principled way, and we obtain good results when we apply this a couple of prediction problems in football, but can be scaled across other sports as well.”
“Two things I believe it helps is it should reduce the amount of time it takes to develop a machine learning model because it works directly with raw data therefore, eliminating the need for time-consuming task of designing and implementing preprocessing stage on the input data, particularly feature engineering.”
“Second, the models provide the basis for advanced statistics that consider the movements and actions of all the players throughout the entire duration of the play. For example, you can measure how difficult a pass attempt is, based on everything that has happened up until the moment the pass is made. This allows the skill of the quarterback and receiver pair to be isolated from the game-state the pass was made in, such as the play selection, defensive strategy, etc.”
Best of luck to Michael and the rest of the team at #SSAC20 this week!
To see what we’re up to at this year’s conference, make sure you’re following us on Twitter at @Sportlogiq.