Interview, Fireside Chat, Conference Presentation
Peter Norvig: We Are Seduced by Our Low-Dimensional Metaphors | AI Podcast Clips
- The speaker prioritizes "trust, validation, and verification" over explainability as the primary goal for neural networks, defining explanations as a tool rather than the solution itself.
- Explainability alone is insufficient for establishing trust, as both humans and machines can provide false explanations that do not reflect reality (e.g., a bank citing collateral when bias was the actual cause).
- A "conversation" is proposed as a superior framework to static explanations, enabling back-and-forth dialogue about decisions, counterfactuals, and necessary changes.
- Detection of bias requires adversarial testing across broad datasets to identify patterns based on protected attributes (religion, skin color) that single-case analysis cannot reveal.
- Current object recognition systems are vulnerable to adversarial attacks; minor pixel perturbations can cause model failures that are "striking" compared to human performance.
- The speaker rejects low-dimensional metaphors (2D maps) for model behavior, arguing instead that data exists in a "million-dimensional space" where stepping slightly off a learned path leads to unpredictable "nowhere's land."
- Addressing model robustness requires a fundamental shift in understanding how models operate within high-dimensional spaces rather than relying on simplified spatial analogies.
- AI systems face a stricter trust barrier than humans, requiring extensive proof of worthiness before receiving any "inkling of trust," whereas humans intuitively trust strangers in shared environments like coffee shops.
- The speaker references naturalist Mark Moffat to highlight human unique ability to coexist with strangers, contrasting it with chimpanzee tribal aggression, suggesting this human social trust is a necessary benchmark for AI.
- The speaker cautions against overemphasizing AI as the primary driver of systemic change, noting that communication technologies enabling data collection and global reach are often the more significant factors.