Title: Towards Cost Efficient Use of Pre-trained Models

Speaker: Prof. Alan Ritter @ Georgia Tech

Time: 14:00 - 15:00, May 20th, 2025

Location: Hybrid

In-person: 85613

Online:  https://hli.skku.edu/InvitedTalk250520-1 

Language: English speech & English slides

Abstract:

Large language models (LLMs) are driving rapid advances in AI, but these breakthroughs come with substantial costs. Training state-of-the-art models demands significant GPU resources for both pretraining and inference, as well as labeled data for post-training. In this talk, I will explore cost-utility tradeoffs that arise across several stages of model development, aiming to inform more efficient decision-making. First, I will examine pretraining-based adaptation, which incurs high computational costs when applied to new domains. Second, I will show that training and distilling large models can offer a cost-effective path to improved performance. Third, I will compare the tradeoffs between supervised fine-tuning and preference-based methods such as Direct Preference Optimization (DPO). Finally, I will present a method for extracting experimental data from scientific tables, enabling automated meta-analyses across thousands of papers on arXiv.org.

Bio:

Alan Ritter is an associate professor in the College of Computing at Georgia Tech.He carried out some of the earliest work on the use of language models to develop chatbots, including training them via end-to-end reinforcement learning.Alan is the recipient of various awards, including an NDSEG Fellowship, NSF CAREER Award, Amazon Research Award, and a Sony Faculty Innovation Award, along with multiple paper awards. His research has garnered media attention from WIRED, TNW, Bloomberg, and VentureBeat.

Title: Cultural Bias and Privacy Protection in Large Language Models

Speaker: Prof. Wei Xu @ Georgia Tech

Time: 15:30 - 16:30, May 20th, 2025

Location: Hybrid

In-person: 85613

Online:  https://hli.skku.edu/InvitedTalk250520-2 

Language: English speech & English slides

Abstract:

In this talk, I will explore two key aspects of large language models (LLMs) and their pre-training data: cultural bias and privacy preservation. First, I will present a systematic study on LLMs' favoritism toward Western culture. Our approach involves curating a dataset CAMEL of over 20,000 cultural items, including food, clothing, individuals, religious sites, and more, to contrast Arabic and Western cultures. We assess cultural biases across multilingual LLMs (e.g., GPT-4, Aya) through natural prompts, story generation, sentiment analysis, and named entity recognition tasks. A key finding suggests heavy reliance on Wikipedia data during pre-training may have contributed to the bias toward Western concepts in non-Western languages.

Second, I will discuss our work on probabilistic reasoning for privacy protection, focusing on estimating the k-anonymity of user-generated text, such as social media posts or exchanges with ChatGPT. We introduce a novel probabilistic reasoning method that leverages LLMs to factorize a joint probability distribution and estimate the number of individuals matching a given set of attributes mentioned in the text. Additionally, we explore text-based disclosure abstraction as a proactive strategy for privacy preservation in PrivacyMirror, an AI-driven privacy protection tool we are developing. Our models detect and rephrase specific self-disclosures into more general terms while preserving their conversational utility. For example, the statement "I'm 16F" can be transformed into "I'm a teenage girl," reducing users' privacy risks while maintaining their intended meaning.

To conclude, I will briefly discuss recent research on the temporal robustness of LLMs, particularly their ability to handle neologisms, and advances in human-AI interactive evaluation.

Bio:

Wei Xu is an Associate Professor in the College of Computing and Machine Learning Center at the Georgia Institute of Technology, where she directs the NLP X Lab. Her research interests are in natural language processing and machine learning, with a focus on Generative AI, robustness and multilinguality of large language models, and interdisciplinary research in AI for science, education, accessibility, and privacy. She is a recipient of the NSF CAREER Award, Faculty Research Awards from Google, Sony, and Criteo, CrowdFlower AI for Everyone Award, Best Paper Awards and Honorable Mentions at COLING'18, ACL’23, and ACL’24. She also received research grants from NIH, DARPA, and IARPA.