Abstract:

Cooperation through safe and trustworthy communication and interaction is fundamental to how human teams accomplish complex tasks. Yet, despite significant–and sometimes revolutionary–advances in AI, we have barely begun to unlock the potential of safe, cooperative AI. This may stem from our limited understanding of how multimodal, large-scale AI models function, the one-sided nature of contemporary, fully-supervised AI approaches, or social concerns about human-AI collaboration. In this talk, I will dive into these layers of inquiry, beginning with a principled exploration of what the embeddings in large-scale foundation models reveal about the underlying problem and data, including new results disentangling sample-size from Bayes error and decision-boundary complexity. I will then introduce the concept of the human collaborator as a “hazy oracle”–a fallible partner rather than an omniscient information source–and establish a framework for modeling human-supplied error during collaboration. Building on these foundational insights, I will conclude with applications of these ideas to foster safe and effective human-AI collaboration in the health sciences.

Zoom Link (if unable to attend in-person): https://upenn.zoom.us/j/99385572884