Our country is deeply divided into two camps.

From coast to coast, people are eager to know the answer to one simple question: Who will come out on top Michael Jordan or LeBron James?

It might seem like a moot point. NBA legend Michael Jordan is now well into retirement while LeBron James is still able to continue building his case with the Los Angeles Lakers. Thanks to the laws of time and space, theres no way to accurately compare their talent in a conclusive way.

Or is there?

AutoStats, a product of Stats Perform, is using artificial intelligence and computer vision to unlock secrets of seasons past and predict seasons future.

The goal of AutoStats is to collect tracking data from every sports video that has ever existed which essentially enables us to travel back in time and compare players and eras in a way that we havent been able to do previously, said Patrick Lucey, chief scientist at Stats Perform. Using this technology, we can start to make the impossible possible.

The implications of these statistics are a real game-changer in the sports world, the effects of which can be seen in betting, team drafting and recruitment, professional commentary, fantasy football and how well your opinions on all-star players hold up.

Sujoy Ganguly, Ph.D.

Director of Computer Vision

I am the director of computer vision, which means I teach computers to watch sports. Specifically, we extract the positions of the players, their limbs and actions directly from the broadcast video you get in your home.

Patrick Lucey, Ph.D.

Chief Scientist

Im the chief scientist, and my role is to set the AI strategy to maximize the value of our deep treasure troves of sports data using AI technology.

Patrick Lucey: AI not only emulates what a human can do, but surpasses what even the best human expert can do. The reason why artificial intelligence has reached this superhuman capability is that it has utilized an enormous amount of data. The more data you have, the better your AI technology will be simple as that.

When it comes to the sheer volume of sports data, no other company has the amount that we have. We cover any sport you can think of, and we capture it at a depth that no other company does.

Sujoy Ganguly: The goal of our team is to create the most in-depth data at the broadest breadth. We do this by extracting player tracking, pose and event data everywhere there is broadcast video. To accomplish this, we have three streams:one that focuses on model development, the second that focuses on the deployment of these models to the cloud, and a third that focuses on implementation at the edge for in-venue deployment.

How does Stats Perform get its data?

Stats Perform collects data through raw video. Its collected via the companys in-venue hardware or snapped up from broadcasts.

Lucey: Well, its like teaching a child how to read. First, they have to learn the alphabet and words before being able to understand a sentence, then onto a paragraph only then they can understand the whole story. Once they have read a lot of books and seen similar stories in the past, then they can actually start to predict how the story will unfold.

Its similar for sport, where we first have to create a sports-specific alphabet and words from which to form sentences that represent gameplay that a computer can understand. Instead of using characters and textual words, we use spatial data and event sequences. From this sports-specific language we have built, we can then get the computer to learn similar gameplay from the data we have, which enables us to predict plays and player motion. The main reason why I believe AI has so much hype around it is that it is the ultimate decision analysis tool every decision and action can be objectively analyzed.

Ganguly: Teaching a machine to interpret sports is a complex and evolving problem. At a high level, we start with a clearly defined question. For example, what is the likelihood that a team will win a game, and how does this depend on the players on that team? Then we ask what information we have: We have results of thousands of games and data about the players who played in those games. From there, we can start the process of conducting experiments and converging to a high-performing model. Generally, this process requires an open and honest conversation about the results of each test and what we have learned.

Ganguly: Many of the challenges we face with machine learning are the same as in other industries, like how we collect and maintain data sets or how we manage training and deployment workloads. However, most companies that work on prediction are doing so on strictly temporal data. In contrast, we have spatial and temporal information. Unlike the autonomous vehicle companies that also deal with spatial-temporal data, we dont control all of the sources of video. This presents unique challenges in data collection but also allows us to use predictive models that allow for noise and are therefore robust.

Different kinds of data

Temporal data is data relating to time and spatial data refers to space. As Ganguly alluded to, combining the two is necessary in the tech behind self-driving cars. This data helps determine whats another moving object, like another car, and whats stationary, say, a tree. For Stats Perform, they data scientists are looking less at a deer in the road, and more how a player moves on the field, and at what speed. The result is the ability to pinpoint the specific motions of a player depending on the context of the game and play, and to anticipate how theyd react in a similar situation.

Lucey: The example I like to talk about is our work in soccer. Soccer is a hard sport to analyze because it is low-scoring, continuous and strategic. As such, the current statistics used, such as possession percentage, number of passes and completion rate, number of corners and tackles, do not correlate with goals scored and who won the match. Our AI-based metrics expected goals, quality of passes and playing styles correlate much higher with goals compared to standard statistics. These AI-metrics simply measure performance better. Using these AI tools, we were able to show how, against incredible odds, underdog Leicester City won the 2015-16 English Premier League title.

Ganguly: There are two significant ways that AI is and will continue to revolutionize sports. Firstly, AI is creating more complex and granular data at an unprecedented scale. For example, with our AutoSTATS technology, we can capture the motions of players in college basketball, where this data was never before available. The other way AI is revolutionizing sport is by allowing people to draw insights from our increasingly in-depth data. Using player tracking data, we can predict the motion of players. This allows us to see how a player will behave on their team after a trade, thereby allowing for better player recruitment.

Isolating a teams formation

Tools like Stats Performsunsupervised clustering method can quickly find a teams formation right down to the frame. When humans attempt to do this, their results fall just a few yards short.

Lucey: Even though we have the most sports data on the planet, to tell the best stories and provide the best analysis and products for our customers, we need even more granular data. Thats why I am so excited about our AutoStats work.

AI has so much hype around it is because it is the ultimate decision analysis tool every decision and action can be objectively analyzed. AI can not only capture data using computer vision and other sensors that couldnt be captured before, but it can help us transform that data into a form that can be used to make decisions. Given how popular sports are around the world and the importance they have on other sectors, theres potential for other industries to directly use the data and technology that we have generated to make future decisions.

Link:

MJ or LeBron Who's the G.O.A.T.? Machine Learning and AI Might Give Us an Answer - Built In Chicago