All Stories

Jonas Olofsson, professor of psychology at Stockholm University, will discuss his research on training language models, similar to ChatGPT, to communicate about perceived odors. He will present his work at 10:30 a.m. Thursday, July 25, in the Sladek Distinguished Science Forum (DSC 163).

Jonas Olofsson Jonas Olofsson Whereas humans learn through interactions with the perceived environment, artificial intelligence (AI) language models (LMs) are only trained on text. This might limit their usefulness as model systems. Smell descriptions are rare in natural language. To represent smells in a human-like way might thus constitute a frontier for LMs.

Prof. Olofsson’s team is investigating how well LMs may perform in psychological smell tasks. They use several human smell perception datasets and test under which conditions LMs can perform similar to humans. They find that simpler LMs sometimes perform better than more sophisticated LMs such as ChatGPT, especially if trained on custom datasets where smell descriptions are common or pre-selected. ChatGPT, on the other hand, performs well on tasks where odor perception is measured by verbal description.

Overall, Prof. Olofsson’s research shows promise for LMs to “talk” about smells, which is of relevance for food, wine, perfume, and other industries.

Sponsoring Department, Office, or Organization:

Psychology Department and Summer Undergraduate Research Experience Program

For more information, contact:

John Kirk: jkirk1@carthage.edu