Keynote – Sameer Singh
summary
Neural nets with millions to billions of parameters achieve near‑human accuracy yet make brittle, counterintuitive errors: a 94% classifier still learns a snow detector; image APIs (Google's Inception) and GPT models produce bizarre outputs. Work uses model‑agnostic local explanations (linear approximations, feature attributions, and anchors), plus paraphrase‑style semantic adversaries and gradient search to find universal triggers that flip labels — e.g., attacks drop accuracy from ~86% to ~29% and generate targeted toxic text on GPT‑style systems.
AI-written summary · about
chapters (8)
- 0:00Detecting ML model bugs
- 7:13Black-box ML model problems
- 13:15Explaining model predictions
- 27:45NLP interpretability tools
- 28:07Bugs with semantic adversaries
- 34:28Universal triggers behavior
- 39:54Call for ML testing standards
- 41:05Q&A on ethics
