Statistical Methods for AI Safety with Industrial Applications

Music City Center 

Companies in every sector of the economy are attempting to leverage the latest advances in artificial intelligence (AI). While this brings exciting opportunities and innovation, quickly incorporating such cutting-edge technology into user-facing systems brings obvious risks. I will summarize my group's recent work on using principled statistical procedures to ensure degrees of reliability in large, black-box AI models (e.g. neural networks, transformers). Applications range from quantifying the uncertainty in visual systems for autonomous driving, to guaranteeing speed-vs-performance tradeoffs for models run on low-resource hardware, to ensuring users cannot prompt a large language model to go beyond behavioral boundaries. All work discussed has been done in collaboration with industrial partners and with contributions from their research scientists.

AI Safety