The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Weakly Supervised Causal Representation Learning with Johann Brehmer - #605
18 snips
Dec 15, 2022 Join Johann Brehmer, a research scientist at Qualcomm AI Research, as he dives into the intriguing world of causal representation learning. He shares insights from his work, emphasizing how high-level causal representations can be identified in weakly supervised environments. They discuss the implications of causality in machine learning, including advancements in autonomous driving and the challenges faced in achieving reliable systems. Brehmer also highlights innovative methodologies and new frameworks for optimizing neural networks and understanding complex causal mechanisms.
AI Snips
Chapters
Transcript
Episode notes
From Particle Physics to Machine Learning
- Johann Brehmer's background is in particle physics, using statistics and machine learning to analyze data from the Large Hadron Collider.
- This led him to discover his passion for the methods themselves, eventually transitioning to machine learning research at Qualcomm.
Causality Research at Qualcomm
- Qualcomm's research spans from applied work like video compression to fundamental research like causality.
- Causality aims to improve AI efficiency by enabling it to reason about changes and robustness, addressing current limitations of standard machine learning.
Causality for Robustness
- Current machine learning systems struggle with brittleness under changing conditions, like sim-to-real transfer.
- Causality offers a framework for reasoning about changes and robustness, potentially addressing this weakness.