Panel Discussion
summary
Panelists debate whether scaling plus architecture wins: larger pretrained models (CLIP/LLMs) empirically fix many classic domain-adaptation traps (e.g., ImageNet waterbirds), yet fail on concrete gaps like stoves in Nepal because web data lacks geographic diversity. Concrete points: scaling helps when tasks reduce to broad recognition; clinical EHRs still need metadata and human-derived features; practical analyses cite 190 Kaggle competitions showing persistent rank-order OOD degradation; theory should move beyond x,y IID to abstractions around next-token prediction, chain-of-thought and controllable, by-construction guarantees.
AI-written summary · about
chapters (9)
- 0:00Panel discussion intro
- 1:03Panelists and their work
- 5:28Scaling in machine learning
- 12:47Universal vs context-sensitive models
- 24:22Misguided research directions?
- 29:15Theory for foundation models
- 41:05Data scaling vs OOD generalization
- 44:04New domain adaptation definitions
- 48:00Closing remarks
related talks
48mFuture of Machine Learning Panel
56mPanel: Scaling AI successfully
41mModels Schm-odels: Why you Should Care about Data-Centric AI
30mEssential Ingredients in Scaling Organizations for ML
33mLive Q&A Session 2
54mKeynote: Can you trust your (large language) model?
30mData-Centric AI Cookbook-let's prep that data
36mWhen the Only Way to Scale up, is to Scale Down