HUMAIN Lab

The HUman-centered MAchine INtelligence (HUMAIN) Lab* has the mission to build controllable machine intelligence that serves humanity safely.

The research in the lab is naturally interdisciplinary. In addition to traditional backbone subjects of machine learning (e.g., deep learning, statistics, optimization), we also gain inspiration from social sciences including economics and psychology about how we could model human preferences and behaviors, and develop rigorous human-subject studies to evaluate machine learning systems and algorithms.

We hope to develop a question-driven instead of a tool-driven research culture in the lab. Thus, our research can be theoretical or empirical, and the technical toolkits we use are diverse, ranging from statistical learning theory to sequential decision-making to causal inference to control theory.

Currently, we focus on three main research pillars in the lab:

1. Understand the foundations of state-of-the-art machine intelligence. The ultimate goal is to use this understanding to build next-generation machine intelligence. For example, we hope to study why the current transformer-based LLMs are so powerful and use the findings as inspiration to design the next-generation LLMs so that they are more factual and reliable.
   • I am looking for students interested in exploring the generalization behaviors of large language models. Reach out if you are interested!
2. Propose principled frameworks for developing human-centered machine learning. Here, focusing on particular machine learning applications that interact with humans (e.g., personalized recommender systems and decision-support systems), we study principled ways to model human preferences and behaviors and incorporate these models into the machine learning pipeline. For more details, please refer to Chapter 1 of my Ph.D. thesis. Recent relevant publication:
   • Personalized Language Modeling from Personalized Human Feedback
     We develop a personalized RLHF framework for fine-tuning personalized language models attuned to a diverse set of human preferences.
3. Evaluate and mediate the societal and economic impacts of large-scale machine learning systems. Our focus here is on machine learning systems—recommender systems and LLMs—that have been deployed to interact with millions of people. In addition to evaluating the impacts of these systems, we hope to develop toolkits to facilitate the implementation of public policies for these systems (e.g., ways to perform data deletion efficiently). Recent relevant publication:
   • Recommender Systems as Dynamical Systems: Interactions with Viewers and Creators
     In societal domains where AI interacts with users, we argue that formal interaction models—–mathematical models that formalize how AI and users influence each other—–can enhance AI design and evaluation practices to achieve a positive societal impact.

Work-life Balance: Finally, it is worth mentioning that the lab values the importance of work-life balance. Given that we are in such a fast-paced field, it takes great effort to keep some time and space for ourselves outside research. To maintain a sustainable working style, we hope to keep the lab small but vibrant.

Prospective Students: We are looking for motivated students who share similar research interests to join the lab. If you are reaching out, please include your CV and why you are interested in joining the lab in your email. If you do not hear back, please feel free to send in a reminder.

*: The lab name is designed by my former officemate Sebastian Caldas and me.