Invited Speakers
Maria Pacheco
Assistant Professor, University of Colorado, Boulder
Maria Pacheco is an Assistant Professor in the Department of Computer Science and a Faculty Fellow in the Institute of Cognitive Science at the University of Colorado Boulder. Before joining CU, she was a Postdoctoral Researcher at Microsoft Research NYC. Maria completed her PhD in Computer Science at Purdue University, and her BSc. in Computer Science and Engineering at the Universidad Simon Bolivar in Caracas, Venezuela, where she was born and raised.
Talk Title: Empowering text-as-data researchers with explainable NLP systems.
NLP systems are everywhere, and the broader research ecosystem is no exception. Researchers in the social sciences, humanities and industry have been making increasing use of NLP technologies to make sense of large textual repositories and answer fundamental questions in their fields of study. When NLP systems are employed as research tools, ensuring their reliability and transparency is incredibly important. Most current systems, while seemingly powerful, are opaque and often fail to satisfy the needs of researchers. In this talk, I will discuss the main challenges that arise when incorporating NLP technologies in text-as-data research, and make the case for explainable, controllable alternatives that can strike a balance between generalization and trustworthiness.
Prasanna Sattigeri
Principal Research Scientist, IBM Research
Prasanna Sattigeri is a Principal Research Scientist at IBM Research AI and the MIT-IBM Watson AI Lab, where his primary focus is on developing reliable AI solutions. His research interests encompass areas such as generative modeling, uncertainty quantification, and learning with limited data. His current projects are focused on the governance and safety of large language models (LLMs), aiming to establish both theoretical frameworks and practical systems that ensure these models are reliable and trustworthy. He has played a significant role in the development of several well-known open-source trustworthy AI toolkits, including AI Fairness 360, AI Explainability 360, and Uncertainty Quantification 360.
Talk Title: LLM Governance Elements: Detection and Alignment
This talk will go over the challenges and opportunities of developing and deploying large language models (LLMs), with a focus on building trustworthy AI systems. We investigate the difficulties of creating a detector library that can label a variety of risks, such as biased or hallucinated outputs. Detectors can be used not only as critical safeguards after deployment, but also for data curation and alignment during the development phase, allowing for effective governance across the LLM lifecycle. In addition, we will go over a few approaches that can help application developers tailor LLM behavior to not only mitigate common harms but also align with specific values or business requirements. Finally, we emphasize the importance of human-centered model evaluation, discussing how explanations and source attribution can improve transparency and trust in LLM applications.
Jieyu Zhao
Assistant Professor, University of Southern California
Jieyu Zhao is an assistant professor of Computer Science Department at University of Southern California. Prior to that, she was an NSF Computing Innovation Fellow at University of Maryland, College Park. Jieyu received her Ph.D. from Computer Science Department, UCLA. Her research interest lies in fairness of ML/NLP models. Her paper got the EMNLP Best Long Paper Award (2017). She was one of the recipients of 2020 Microsoft PhD Fellowship and has been selected to participate in 2021 Rising Stars in EECS workshop. Her research has been covered by news media such as Wires, The Daily Mail and so on. She was invited by UN-WOMEN Beijing on a panel discussion about gender equality and social responsibility.
Talk Title: Building Accountable NLP Models for Social Good
The rapid advancement of natural language processing (NLP) technologies has unlocked a myriad of possibilities for positive societal impact, ranging from enhancing accessibility and communication to supporting disaster response and public health initiatives. However, the deployment of these technologies also raises critical concerns regarding accountability, fairness, transparency, and ethical use. In this talk, I will discuss our efforts for auditing NLP models, detecting and mitigating biases, and understanding how LLMs make decisions. We hope to open the conversation to foster a community-wide effort towards more accountable and inclusive NLP practices.
Ahmad Beirami
Research Scientist, Google Research
Ahmad Beirami is a research scientist at Google Research, co-leading a research team on building safe, helpful, and scalable generative language models. At Meta AI, he led research to power the next generation of virtual digital assistants with AR/VR capabilities through robust generative language modeling. At Electronic Arts, he led the AI agent research program for automated playtesting of video games and cooperative reinforcement learning. Before moving to industry in 2018, he held a joint postdoctoral fellow position at Harvard & MIT, focused on problems in the intersection of core machine learning and information theory. He is the recipient of the 2015 Sigma Xi Best PhD Thesis Award from Georgia Tech.Talk Title: Language Model Alignment: A Theoretical View
The goal of the language model alignment (post-training) process is to draw samples from an aligned distribution that improves a reward (e.g., make the generation safer) but does not perturb much from the base model. A simple baseline for this task is best-of-N, where N responses are drawn from the base model, ranked based on a reward, and the highest ranking one is selected. More sophisticated techniques generally solve a KL-regularized reinforcement learning (RL) problem with the goal of maximizing expected reward subject to a KL divergence constraint between the aligned model and the base model. An alignment technique is preferred if its reward-KL tradeoff curve dominates other techniques. In this talk, we give an overview of language model alignment and give an understanding of known results in this space through simplified examples. We also present a new modular alignment technique, called controlled decoding, which solves the KL-regularized RL problem while keeping the base model frozen through learning a prefix scorer, offering inference-time configurability. Finally, we also shed light on the remarkable performance of best-of-N in terms of achieving competitive or even better reward-KL tradeoffs when compared to state-of-the-art alignment baselines.
