The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Disentangled Representations & Google Research Football with Olivier Bachem - TWIML Talk #293
Aug 22, 2019
Olivier Bachem, a research scientist at Google AI's Brain team, dives into exciting advancements in reinforcement learning. He shares insights on Google Research Football, a unique environment designed for AI training that outshines traditional platforms like OpenAI Gym. Olivier discusses the intricacies of disentangled representations in high-dimensional data and the innovative challenges of developing soccer simulations. He also touches on recent updates, emphasizing collaborative learning and realistic gameplay dynamics in this pioneering research project.
AI Snips
Chapters
Transcript
Episode notes
Disentangled Representations
- Olivier Bachem has worked on disentangled representations, aiming to capture underlying factors of variation in high-dimensional data like images.
- This involves isolating properties like object position, size, and color into separate representational elements.
Challenges of Disentanglement
- Disentangled representations are challenging to evaluate and compare due to the lack of ground truth labels during training.
- Model selection is difficult, as even with identical hyperparameters, results vary significantly, making it hard to identify optimal models without label access.
Project Origins
- Google Research Football started as a team idea at Google Brain to apply reinforcement learning to soccer video games.
- The project leveraged the open-source Gameplay Football as a foundation, offering flexibility for modification and research.