2024 India Retreat | Track 2+ 3, Session 4

Summary

The engineering director of AML discusses the team's journey in creating customer experiences using machine learning. From the inception in 2019 to future plans, the talk covers partnerships with Indian colleagues and other business units, the evolution of products like Instat and Huddle, and the integration of various data types. The director also highlights machine learning's role in enhancing tactical generation, streamlining fan experiences, and improving data quality. The presentation concludes with a Q&A session addressing collaboration strategies, market trends, and future priorities for AML.

Takeaways

• 📈 The Applied Machine Learning (AML) team at Huddle has been instrumental in creating customer experiences and breaking into new markets since 2019.
• 🤖 AML leverages machine learning to optimize and automate various aspects of sports data analysis, such as event detection, tactical camera positioning, and player tracking.
• 🏀 Products like Instat and Huddle focus on different types of data, from event-based analysis to tracking and physiological metrics, to cater to the needs of professional sports teams.
• 🔍 AML's approach to data involves a spectrum of complexity, from simple events to sophisticated game insights, which helps in understanding and predicting game dynamics.
• 🔄 The iterative feedback loop of machine learning is emphasized, where data annotation, model training, deployment, and correction are continuously refined to improve accuracy.
• 📊 AML has made significant strides in automating tagging processes, reducing manual effort, and increasing the efficiency of data generation for sports analysis.
• 📹 The team is working on projects like 'Ad Trigger' for ad placement in live streams and 'Broadcast Tracking' to estimate player positions for physical metrics, even when not in view.
• 🌐 AML collaborates closely with other departments, treating them as partners rather than customers, to understand and meet the needs of end-users more effectively.
• 🚀 Upcoming priorities for AML include expanding automation to player-level events, generating physical data to replace traditional combines, and improving the setup process for cameras.
• 🌟 The script highlights the importance of keeping up with market trends and customer expectations, such as real-time data and sophisticated insights, to stay competitive.
• 🔮 Looking ahead, AML aims to improve data quality, reduce tagging costs, and expand into new markets and sports, focusing on a human-in-the-loop approach to enhance machine learning models.

Q & A

What is the role of the Applied Machine Learning (AML) team at Huddle?

The AML team at Huddle is responsible for creating experiences for customers, partnering with colleagues in India and other business units, and looking at the future of AML to understand how it impacts products and services.

How does AML use machine learning in its products?

AML uses machine learning to create insights from video and data that help coaches and athletes grow. They analyze different types of data such as event data, scoring data, tracking data, physiological, and physical data to enhance their products.

What is an example of an event-based product in AML's portfolio?

Instat is an example of an event-based product used by professional basketball and ice hockey teams. It analyzes trends and data across games, breaking down footage, and identifying areas for player improvement.

How does Huddle Studio differ from event-based data products?

Huddle Studio creates graphics for coaches on video to aid team development. It uses data to determine camera positions and distances between players, allowing for the creation of informative graphics that remain consistent as the camera moves.

What is the purpose of the 'Ad Trigger' project in AML?

The 'Ad Trigger' project aims to identify optimal placement for advertisements during live video streams of games. It finds the start and end of plays and places ads during dead periods to enhance the viewing experience for fans.

What are the five different types of data that AML works with?

AML works with event data, scoring data, tracking data, physiological data, and physical metrics. Each type serves different purposes and is used in various ways across their products.

Can you explain the concept of 'Player Hallucinations' mentioned in the script?

'Player Hallucinations' is a term used to describe the process of estimating the location of players on the pitch when they are not in the camera's view. This is important for calculating physical metrics such as total distance or sprint speed.

What is the significance of the 'Supercharge Assist' project in AML's history?

The 'Supercharge Assist' project was significant as it aimed to streamline the substitution annotation process, making it more efficient and reducing the manual effort required. It also automated aspects of American football formations prediction.

How does AML plan to improve its data generation and machine learning models?

AML plans to improve data generation by integrating humans into the training and prediction process, known as 'human in the loop'. They aim to reduce the retraining time of models from years to weeks, increasing the iteration speed and accuracy of predictions.

What are some of the future trends in machine learning that AML is keeping an eye on?

AML is particularly interested in the trend of AI alignment, which focuses on ensuring that the goals of artificial superintelligence align with human values and goals to create a better world.

What are AML's top priorities for the next 12 to 24 months?

AML's top priorities include automating team-level events for Focus Flex cameras, generating physical data for American football to replace combines, extending automation into player-level events, and improving the setup process and UI for better user experience.