Talk Abstract
The artist Harold Cohen was among the first to write art-making code (the AARON system) When he died in 2016 he left a huge archive of artworks, code, writing, correspondence and ephemera. I have been working on AI methods to organize the archive to support scholars and museum-goers, including a question-answering system based on the Claude LLM. Provenance and factuality are essential, so retrieval augmented generation (RAG) is a given, but it is not sufficient. Scholarly corpora have hidden structure without which LLMs conflate events, neglect influences, ignore coincidences, and often create Frankenstein's monsters from quite different concepts. Performance improves dramatically when the LLM is given a compact model of the subject along with RAG results. However, curating models is time-consuming, so I have developed AI tools to help scholars build models of the hidden structure of their digital archives. This talk features a lot of fine art and it is open to anyone interested in AI-enhanced scholarship in the arts and humanities.
Biography
Paul Cohen is a Professor of Computer Science at Pitt. He has served as the Founding Dean of the School of Computing and Information, as a DARPA program manager, as Trustee of the Harold Cohen Trust and as CEO of Causerie.AI, a DARPA-supported startup. Cohen is a Fellow of the Association for the Advancement of Artificial Intelligence.
Personal Website
Location
Sennott Square, Room 5317
Date
Friday, March 6 at 12:30 to 1:30 p.m.