Amy Zhang is a researcher in artificial intelligence, focusing on human-centered AI, reinforcement learning, and interpretability. She is an assistant professor at the University of Washington and a research scientist at Meta AI (formerly Facebook AI Research). Her work aims to make AI systems more transparent, interactive, and aligned with human values, often combining machine learning with insights from cognitive science and human-computer interaction.
Zhang received her Ph.D. in computer science from the University of California, Berkeley, where she was advised by Pieter Abbeel. Her doctoral research developed methods for learning from human feedback and for making reinforcement learning more sample-efficient. She also spent time as a visiting researcher at Google Brain and at the Montreal Institute for Learning Algorithms (MILA). Before her Ph.D., she earned a bachelor's degree in computer science from the University of Waterloo.
Reinforcement Learning and Human Feedback
Zhang's early work focused on reinforcement learning, particularly on how agents can learn from sparse rewards and from demonstrations. She contributed to the development of algorithms that combine imitation learning with reward shaping, enabling agents to learn complex behaviors from a small number of human examples. One notable project, "Learning by Playing," introduced a method for solving sparse-reward tasks by using unsupervised exploration to learn skills that can be reused for downstream tasks.
A central theme in her research is the use of human feedback to guide AI training. She has worked on techniques for preference-based reinforcement learning, where a model learns from pairwise comparisons of trajectories rather than explicit rewards. This approach is particularly useful in domains where designing a reward function is difficult, such as robotics or dialogue systems. Her work in this area has been cited widely and has influenced subsequent research on RLHF, a key component in training modern large language models.
Interpretability and Transparency
Zhang has also made significant contributions to interpretability in machine learning. She developed methods to visualize and understand what neural networks have learned, particularly in the context of reinforcement learning agents. Her work on "saliency maps" and "concept activation vectors" helps researchers identify which features of an input are most influential in a model's decision. This line of research is crucial for building trust in AI systems, especially in high-stakes applications like healthcare or autonomous driving.
In a widely cited paper, Zhang and colleagues introduced "TCAV" (Testing with Concept Activation Vectors), a technique that quantifies how much a concept (e.g., "striped" or "spotted") contributes to a model's prediction. This method allows users to probe a model's reasoning in human-understandable terms, moving beyond simple feature attribution. The work has been adopted by industry teams at Google DeepMind and OpenAI for model auditing.
Human-Centered AI
Zhang's research agenda is explicitly human-centered: she designs AI systems that can interact with people naturally and learn from their feedback. She has explored how to make AI assistants more helpful by allowing users to correct mistakes through natural language or by demonstrating desired behavior. This includes work on interactive learning, where a model improves over time based on user input, and on aligning AI objectives with human preferences.
One of her projects, "Reward Learning from Human Preferences," demonstrated that a robot can learn to perform tasks by receiving feedback from non-expert users. In a user study, participants taught a simulated robot to navigate and manipulate objects by providing simple yes/no feedback. The results showed that the robot could learn effectively with minimal human effort, suggesting a practical path toward customizable AI.
Academic Career and Teaching
Zhang joined the University of Washington in 2021 as an assistant professor in the Paul G. Allen School of Computer Science & Engineering. She leads the Human-Centered AI Lab, where she mentors graduate students and postdocs. Her teaching includes courses on reinforcement learning and on fairness, accountability, and transparency in machine learning. She has received several teaching awards, including the UW College of Engineering Outstanding Teaching Award in 2023.
At MIT CSAIL and Stanford AI Lab, she has given invited talks on her research. She is also an affiliated faculty with the UW Center for an Informed Public, where she studies the societal impacts of AI, including misinformation and algorithmic bias.
Industry Experience
Before her academic appointment, Zhang worked as a research scientist at Meta AI from 2020 to 2021. At Meta, she applied her expertise in reinforcement learning to improve recommendation systems and to develop tools for detecting harmful content. She also collaborated with teams at Anthropic on safety research, contributing to early work on scalable oversight.
Zhang has also been a research intern at Google Brain and at Berkeley AI Research (BAIR). These experiences gave her a broad perspective on how AI research translates into products and policy.
Awards and Recognition
Zhang has received several honors for her work. In 2022, she was named a Rising Star in AI by the AI2050 initiative. In 2023, she received the NSF CAREER Award for her project on "Interactive and Interpretable Reinforcement Learning." Her papers have won best paper awards at major conferences, including the International Conference on Machine Learning (ICML) and the Conference on Neural Information Processing Systems (NeurIPS). She is also a recipient of the Sloan Research Fellowship in Computer Science (2024).
Selected Publications
Zhang has published over 40 papers in top-tier venues. Some of her most cited works include:
- "TCAV: Relative Concept Importance Testing" (ICML 2018) - introduced a method for concept-based interpretability.
- "Learning by Playing: Solving Sparse Reward Tasks from Scratch" (ICML 2018) - proposed a curriculum of self-play for exploration.
- "Reward Learning from Human Preferences" (NeurIPS 2019) - demonstrated practical preference-based RL.
- "Interactive Learning from Policy-Dependent Human Feedback" (ICML 2020) - studied how feedback changes as the agent improves.
Her research has been supported by grants from the National Science Foundation, the Office of Naval Research, and private foundations.
Future Directions
Zhang's current work focuses on making large language models more interpretable and controllable. She is investigating how to elicit honest uncertainty from models and how to enable users to steer model behavior without retraining. She is also interested in multi-agent reinforcement learning, where multiple AI systems interact, and in ensuring that such systems remain safe and aligned with human goals.
As of 2025, Zhang continues to teach and lead research at the University of Washington. Her contributions have helped shape the emerging field of human-centered AI, bridging the gap between technical machine learning and human needs.