The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Evolving AI Systems Gracefully with Stefano Soatto - #502
Jul 19, 2021
Stefano Soatto, VP of AI Application Science at AWS and a professor at UCLA, dives into the fascinating world of Graceful AI. He discusses the challenges of evolving AI in real-world applications while avoiding the pitfalls of constant retraining. Topics include the critical timing of regularization in deep learning, the parallels between model compression and material science, and the intricacies of model reliability. Stefano also unpacks innovations like focal distillation and their potential to enhance lifelong learning in AI systems.
AI Snips
Chapters
Transcript
Episode notes
True Intelligence
- Building truly intelligent autonomous systems, like a dog, is more challenging than creating a chess-playing program.
- A dog's intelligence in navigating the real world surpasses a computer's ability in a specific domain.
Negative Effects of Retraining
- Constantly retraining machine learning models in production, while necessary for handling data drift, can introduce negative effects.
- These effects raise questions about the ideal approach to model evolution in live systems.
Customer Obsessed Science
- Customers hesitated to update to better-performing models due to regressions, or new errors the models made.
- This highlighted the mismatch between academic metrics (average error) and customer needs.