Artificial Intelligence Compliance - A Transdisciplinary Challenge

Register: https://rensselaer.webex.com/weblink/register/r6cf0b2235a7e5e9023daa4c10160062d

TALK 1 (WSTNet Distinguished Talk)

Artificial Intelligence Compliance - A Transdisciplinary Challenge
Prof. Steffen Staab | University of Stuttgart, Germany, and University of Southampton, UK

Abstract

Compliance means conforming to a set of rules, requirements, or standards that apply to an organization or individual. AI compliance is often discussed as a technical challenge to be achieved algorithmically. In contrast, we draw on other compliance challenges, e.g., financial compliance, and sketch a landscape in which many stakeholders from diverse backgrounds and disciplines must collaborate to achieve AI compliance throughout the lifecycle of an AI system.

Speaker bio

Professor Staab is a professor of Analytic Computing and heads the Institute for Artificial Intelligence, the Cluster of Excellence for „Data-integrated simulation science“ (SimTech), and the Research Initiative „Interchange Forum for Reflecting on Intelligent Systems“ (IRIS) at the University of Stuttgart, Germany. He also holds a chair for Web and Computer Science at the University of Southampton, UK. Professor Staab studied computer science and computational linguistics in Erlangen (Germany), Philadelphia (USA), and Freiburg (Germany). From 1998 to 2004, Steffen worked as a project lead and lecturer at KIT in Germany, and from 2004 to 2020 as a professor of database and information systems at the University of Koblenz in Germany. At Koblenz, he founded the Institute for Web Science and Technologies (WeST). He is a fellow of the ACM, ELLIS, EurAI, and AAIA. In his research career, he has avoided almost all of the good advice he now gives to his team members. Such advice includes focusing on research (vs. the company) or on only one or two research areas (vs. considering ontologies, knowledge graphs, Web science, data management, machine learning, simulation, and more). 

--------------------------------------------------

TALK 2 (WSTNet International Seminar)

STEREOTYPEMINER: Automatically Mining Stereotypical Semantic Axes in LLMs
Farane Jalali Farahani | Institute for Artificial Intelligence, University of Stuttgart, Germany

Abstract

Stereotypes in large language models (LLMs) have been studied along a small set of predefined axes drawn from social psychology, such as warmth and competence. In this work, we introduce STEREOTYPEMINER, a framework that automatically discovers stereotypical semantic axes directly from LLM internal representations. STEREOTYPEMINER probes attention head activations to represent semantic axes as vectors, projects social group mentions onto them, and identifies stereotypical semantic axes as those along which social groups spread out significantly more than random nouns. We validate STEREOTYPEMINER against psychological surveys and demonstrate its key advantages: it outperforms layer-averaging baselines, reveals that stereotypical associations are concentrated in specific attention heads, and discovers novel stereotypical axes beyond those prescribed by social psychology theory. Our results suggest that LLMs encode a richer set of stereotypical associations than previously recognized.

Speaker bio

Farane Jalali Farahani is currently pursuing her PhD studies at the Institute for Artificial Intelligence at the University of Stuttgart, Germany. She holds a background in computer science. Her research interests span natural language processing and computational social science.

Link to register