flâneur

Red teams. Training AI systems to avoid… | by Paul Christiano | AI Alignment

ai-alignment.com · 3,478 words · saved by 2 readers

Training AI systems to avoid catastrophic errors — without causing catastrophes.

Machine Learning Artificial Intelligence Red teams Paul Christiano 14 min read · May 28, 2016 -- 5 Listen Share To build more robust machine learning systems, we could use “red teams” who search for inputs that cause catastrophic behavior. I current believe this is the most realistic long-term approach to building highly-reliable systems. A “red team” might try to find an image of a black couple that a classifier labels gorillas , or a simulated situation in which a household robot cooks the family cat. These inputs would then be incorporated into the training and development process. If the r

saved by

related reading