Center for Uncertainty Studies Blog

An illustration showing how LLMs deal with presuppositions, used in the talk by Sina Zarrieß.

On June 17, 2024, the second CeUS Uncertainty Research Afternoon took place at Bielefeld University. The event featured talks by computational linguist Sina Zarrieß, bioinformatician Martin Lewinski, and mathematical economist Maren Diane Schmeck.

Sina Zarrieß discussed uncertainty in large language models (LLMs, e.g., ChatGPT) in three linguistic dimensions: semantics, pragmatics, and interaction. She demonstrated that in each dimension, the analyzed LLMs lack the ability to make their own uncertainties transparent, even in examples classified as explainable AI which add certain explanations to their utterances. Consequently, the concept of uncertainty is not present in the analyzed transformer language models. The group discussion connected these limitations with other issues concerning LLMs and AI in general.

In the next talk, “It Takes Two to Tango – RNA and Protein Uncertainties from the Wet Lab to Bioinformatics,” Martin Lewinski continued the interdisciplinary discussion, combining biology and informatics. He and his colleagues grow and analyze thale cress (Arabidopsis thaliana), a plant with a well-annotated and small genome suitable as a model for crop plants, to find out more about RNA-binding proteins (RBPs). Using the iCLIP method, biological, technical, and bioinformatic uncertainties arise in the research process. Lewinski's role, primarily in the bioinformatics realm, involves identifying and minimizing these uncertainties. The debate raised further questions: what can't we capture when "fishing for uncertainties"?

The final presentation of the afternoon was by mathematical economist Maren Diane Schmeck, who discussed her research, “On Uncertainty in Time Evolution” Her research starts from the premise that calendar time might not be the appropriate measure of time in financial markets.  Schmeck demonstrates that mathematical models incorporating changes in the clock-speed can account for several empirical stylized facts on financial markets, while still relying on standard Brownian motions to capture stochastic influences on the market. These models are used to estimate market activity, for example, in the electricity market, in ‘normal’ times and in times of crisis, and thereby can provide valuable insights into trading behavior under uncertainty.

The Uncertainty Research Afternoon once again proved to be a challenging yet highly inspirational interdisciplinary event with innovative research and lively debate. Join the conversation at our next Uncertainty Talk with Maida Kosatica (Duisburg-Essen) on July 8. More information can be found on the CeUS website.