This is what DeepMind just did to Football with AI...

Episode Overview

• The episode introduces "TacticAI," a new AI assistant for football tactics developed by Petar Veličković and his team at DeepMind in collaboration with Liverpool Football Club.
• It explains how TacticAI uses Geometric Deep Learning (GDL) and Graph Attention Networks (GATs) to analyze player positions and suggest optimal strategies, particularly for corner kicks.
• The system combines predictive models (what will happen) and generative models (how to make something happen) to provide actionable insights for coaches.
• A short clip from an interview with Petar Veličković provides context on the project's goals and its potential as a practical tool for football clubs.

Key Concepts

• TacticAI: A full AI system with predictive and generative models to analyze football plays. It can assess past plays, predict outcomes, and suggest adjustments to make desired outcomes more likely.
• Geometric Deep Learning (GDL): The underlying machine learning paradigm that leverages symmetries and relational structures in data. In this context, it models players as nodes in a graph to capture interactions.
• Graph Attention Networks (GATs): The specific architecture used by TacticAI. It was co-invented by Petar Veličković and is well-suited for modeling the relationships and permutations between players on the field.
• Corner Kick Optimization: The primary application discussed for TacticAI. The model analyzes corner kick setups and generates alternative player positions that are indistinguishable from real tactics and favored by experts 90% of the time over existing ones.
• Human-AI Collaboration: TacticAI is framed as an assistant to augment the capabilities of human coaches by automating tedious analysis and inspiring creative, data-driven tactical refinements.

Quotes

• At 01:00 - "favored over existing tactics 90% of the time." - The speaker highlights the remarkable success rate of TacticAI's suggestions when evaluated by human football experts.
• At 03:30 - "to study tactics in football, which was a project that we have done together with Liverpool Football Club." - Petar Veličković introduces the project and its collaboration with a major professional sports team.
• At 04:12 - "we actually believe we have an AI football tactics assistant on our hands, because TacticAI is capable not only to predict what will happen... but it's also capable of giving you recommendations." - Veličković explains why TacticAI is a significant step forward, moving from simple prediction to actionable, generative suggestions.

Takeaways

• AI is becoming a powerful assistant in sports analytics, capable of suggesting novel and effective strategies that can augment a coach's expertise.
• Geometric Deep Learning and Graph Neural Networks are highly effective for modeling complex, relational systems like team sports, as they can capture player interactions and symmetries.
• The combination of predictive and generative AI allows models to not only analyze what has happened but also to explore "what if" scenarios and suggest improvements.
• The high approval rate (90%) from human experts indicates that AI-generated tactics can be both plausible and superior to existing human strategies, freeing up coaches to focus on creative, high-level decision-making.