Home Call For Papers Organizers Keynote Accepted Papers Schedule

Attending Program

2-6pm, Room 120, August 25th, 2024, Centre de Convencions Internacional de Barcelona, Barcelona, Spain

2:00 – 2:10 pm Opening Remarks
2:10 – 2:40 pm Keynote Talk 1: Toward Ethical AI for the Neglected Tails
Dr. James Caverlee, Texas A&M University
Abstract: In this talk, I will share some of our recent work centered on fairness and bias with an eye toward the “neglected tails” in applications like recommendation, LLMs, vision-language models, and speech systems. For example, these systems often demonstrate strong performance on popular concepts (or items or users), but in many cases there is a gap in the treatment of rare (or tail) concepts. Can we bridge this gap? These and similar questions are motivated by the Rawlsian max-min principle to design systems that maximize the least well-off (or “tails” of the distribution).
2:40 – 3:40 pm Accepted Paper Talks 1
  • [Paper ID: 4] Practical Fairness: An Evaluation of Bias Mitigation Strategies for NLP Applications.
    Bryce King, Leidos
  • [Paper ID: 5] Fair Data Generation via Score-based Diffusion Model.
    Minglai Shao, Tianjin University
  • [Paper ID: 6] Matchings, Predictions and Counterfactual Harm in Refugee Resettlement Processes.
    Suhas Thejaswi, Max Planck Institute for Software Systems
  • [Paper ID: 8] POSIT: Promotion of Semantic Item Tail via Adversarial Learning.
    Ding Tong, Netflix Research
  • [Paper ID: 9] AI Safety in Practice: Enhancing Adversarial Robustness in Multimodal Image Captioning.
    Maisha Binte Rashid, Baylor University
3:40 – 4:10 pm Keynote Talk 2: Trustworthy LLMs: Detection and Red-Teaming
Manish Nagireddy, IBM Research
Abstract: Trustworthy AI is paramount to the responsible use of AI systems. Despite the rising popularity of large language models (LLMs), their generative nature amplifies existing harms related to trust (such as fairness, robustness, transparency, etc.) and reveals new dangers (such as hallucinations, toxicity, etc.). I will first go through a catalog of harms and delve more deeply into how such harms can be automatically measured with detectors. Notably, these detectors can be applied throughout the LLM lifecycle (from filters on pre-training data to reward models during alignment to guardrails after deployment). Then, I will go through an example of developing a benchmark to capture a unique harm that was discovered via interactive probing. Next, I will combine both ideas to describe the development of a nuanced detector. Finally, I will end with future thoughts on the need for dynamic and participatory evaluation practices (such as red-teaming) and next steps for more trustworthy systems.
4:10 – 4:25 pm Coffee Break
4:25 – 4:55 pm Keynote Talk 3
Dr. Tyler Derr, Vanderbilt University
4:55 – 5:55 pm Accepted Paper Talks 2
  • [Paper ID: 11] Source Echo Chamber: Exploring the Escalation of Source Bias in User, Data, and Recommender System Feedback Loop.
    Sunhao Dai, Renmin University of China
  • [Paper ID: 12] OxonFair: A Flexible Toolkit for Algorithmic Fairness.
    Eoin Delaney, University of Oxford
  • [Paper ID: 15] Enhancing Model Fairness and Accuracy with Similarity Networks: A Methodological Approach.
    Samira Maghool, University of Milan
  • [Paper ID: 16] A Semidefinite Relaxation Approach for Fair Graph Clustering.
    Sina Baharlouei, eBay
5:55 – 6:00 pm Closing