I am Zhiyu Lin (林之雨)

I am Zhiyu Lin, a Ph.D. candidate from Georgia Institute of Technology, advised by Professor Mark Riedl. My distinct combination of engineering expertise, leadership skills, and experience in studying humans and AIs both quantitatively and qualitatively to build algorithms and frameworks with human values in the center of the stage makes me distinct from other researchers. They enable me to perceive how value-centered AI can be beneficial as well as the deficiencies of the existing methods used in cutting-edge AI and ML technologies beyond the buzz, through the lens of human-centered AI and computational co-creativity.

My Research Goals

I have expanded my understanding of AI and humans during my career as a researcher and engineer, particularly when they work toward a common creative objective. In contrast to the widespread idea that AI will eventually replace humans, my research has shown the potential of AI agents that can collaborate and learn to prioritize human values in order to outperform both human and computational creativity alone. Following these beliefs, I have identified mixed-initiative co-creative systems as my next steps, where humans and AI can work together while respecting each other's goals and values. I will continue to research how computational creativity can cooperate with human creators by interacting with each other and updating how they work with their counterparts while being explainable and trustworthy, ultimately empowering this assemblage beyond humans or AI alone.

How did I discover my interest?

I was born and raised in Shanghai, China, a city well recognized for fusing numerous cultures to create its own. Growing up as a child in the 2000s, I was surrounded by exciting developments in science and technology and was extremely interested in technical topics, and was fortunate enough to be among the first group of children in my city to receive formal Computer Science (CS) education on algorithms and data structures at the age of nine. I naturally advanced toward becoming a good software engineer because I have a strong interest in computer science. I started by winning the regional high school champion title, which is awarded to the top 0.5% of qualified competitors, in the Olympiad in Information. Then, I went on to earn my bachelor's degree in software engineering from Shanghai Jiao Tong University and my master's degree in software engineering from Carnegie Mellon University before beginning my studies for a Ph.D. in computer science at Georgia Institute of Technology. My ability to thoroughly investigate prospective obstacles, critically assess them, make plans, and effectively communicate and manage a team has been shaped by my strong understanding and experience of how to use computers to build usable engineering products both from an engineering and management perspective.

I worked as a full-stack developer during my internship as a software engineering student at the local IT firm yeah-info.com, which specializes in web apps that deliver content such as videos and music to consumers. My workday begins, like that of every other engineer in the organization, by gathering requirements from the planning and designing team, creating engineering artifacts that comply with their specifications, and sending the finished artifacts to the testing team in the hopes that we won't experience bounce-back or need to redo the engineering due to constantly changing and vague requirements or bugs discovered by the testing team. I soon discovered that creativity, problem-solving skills, and the capacity to create what I desire were not appreciated. Moreover, computers were unable to communicate with engineers, collaborators, and most crucially, stakeholders, to define the requirements of the artifacts. As a result, the performance of the engineering team dropped with these bounce-backs. I began to wonder: Can computers be more conscious of what they are doing? How can future AI tools - rather than merely being passive, simple tools - understand creators' intentions, work with them, and uphold their values?

This, together with my extensive experience in software engineering principles, sparked a great interest in investigations into co-creative and mixed-initiative AI that fosters and encourages human creativity.

About my research

As a Ph.D. student, I worked on my interests and developed my expertise with two main focuses: How AI, especially neural generators, should promote creative expressions of human creators to enhance their capability in fulfilling their goals, and how AI agents, especially Reinforcement Learning (RL) ones, should be aware of its deficiency and learn from human to improve themselves towards being a better collaborator and co-creator.

On one hand, As part of my exploration into procedurally generating content with humans in mind, I studied neural, data-driven models that are well-controllable by content creators with high-level concepts from them. First off, while working with an undergraduate student, I oversaw the GenerationMania project, where we developed an experimental human-aware content generation system that resulted in the publication "GenerationMania: Learning to Semantically Choreograph," which was nominated for the Best Student Paper award at the 15th AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment (AIIDE 2019). By extracting semantic data from the music, fusing it with the authoring goals, and using neural networks to predict which sounds in the music are to be played by the player and which actions players should carry out to recreate the music, I contributed techniques in this work that create a game stage from both music and creative goals from game stage creators.

The generation system developed in this study is the first of its kind because it not only completes the task but also makes it possible to incorporate design ideas directly from human users, which intrigued my curiosity in looking into this subject further.

In light of these findings, we began to consider how we may concentrate on utilizing high-level user intentions. This led to the Plug-and-Blend project, which resulted in the publication “Plug-and-blend: a framework for plug-and-play controllable story generation with sketches”, for which the engineering artifacts used in this research were nominated by AIIDE 2021 for Best Artifact. This work focuses on the domain of storytelling with an awareness of human intents.

In order to create a backbone generator that comprehends the high-level intentions of human authors while still being relevant to this day, we developed a plug-and-play framework over existing GPT-based language models that knows how to incorporate loose or fine-grained ideas in the form of "topic controls" more familiar to human authors than prompts. Our optimism that we can develop AI that recognizes the importance of human values and works in collaboration with them was strengthened by the fact that human subject study participants rated our framework statistically better at generating stories that follow creators' goals.

On the other hand, I researched AI agents that detect novelties in their environment and adapt to them by gathering data, including input from people. I participated as an engineering lead and the mentor of two undergraduate students on a multi-disciplinary, multi-department investigation team that looked into how RL agents should recognize “novelties”, such as the appearance of unanticipated objects or shifting world dynamics, and react. In order to increase agent performance following novelty emergence, we practiced novelty detection and adaption techniques that enhance agent performance in activities like crafting in Minecraft. We also looked at the systematic characterization of novelties as changes in models of the world. These investigations led to the publication "Novgrid: A flexible grid world for evaluating agent response to novelty", to which I contributed, and other papers that are currently being reviewed or written. These investigations gave me insight into the potential of AI agents that actively explore and alter themselves in order to improve themselves.

Together with Dr. Brent Harrison, then a postdoctoral researcher and now an assistant professor at the University of Kentucky, we created an RL agent that learns to solicit human feedback, in the form of approval or disapproval, only when necessary to improve performance in virtual three-dimensional environments. We did this by modeling the agent's confidence in its own policies and the consistency of human input. This project led to the paper “Explore, exploit or listen: Combining human feedback and policy model to speed up deep reinforcement learning in 3d worlds”, and suggested to me the potential of a collaborative, intelligent agent that is aware of its weaknesses and actively centers itself on the value of the people it works with.

To bring the best of generative and human-value-aware AI together, I led the Creative Wand project while I mentor four undergraduate students, which examined how an AI should support mixed-initiative co-creative experiences, enabling the creation of contents that are superior to what either human or AI could produce on their own. We focus on communicating ideas and intents between the user and the agent, and produced the publication "Creative Wand: A System to Study Effects of Communications in Co-creative Settings" at AIIDE 2022, as well as work that is under review and in progress. We create a flexible, modular framework for experimental mixed-initiative co-creative systems, and then use this framework to research the realm of communications outside of prompts and control codes common in current controllable generation systems. While we explore the effectiveness of systems with more sorts of communication in areas like storytelling and game worlds using this framework, we discovered both quantitatively and qualitatively needs from creators for the AI system to be tailored to their requirements and ideals. This further emphasized the significance of value-centered AI systems.

My academic services

As a researcher, I am aware that our primary duties include serving the academic community and distributing knowledge as a mentor. I am an active member of the academic community, participating in program committees and reviewing papers for a variety of conferences and workshops, particularly those that highlight my expertise in comprehending computational creativity and human awareness of AI.

Within the institution, I was a teaching assistant for the Game AI course, responsible for both grading and question-answering, and I personally advised 8 students on research projects throughout my Ph.D. program, while co-advising more students with other collaborators. Being the "big brother" at work because I enjoy exchanging new research ideas with everyone and providing inspiration when needed, I discovered that respect for human values extends beyond AI to my friends, team members, and colleagues, with diversity in their cultural, educational, and professional backgrounds.

My life

When I'm not at work, I enjoy jogging, cycling, visiting places, techno/electro music. I usually act as @xxbidiao (lit. Mini Pidgeot) on the web. 

I enjoy playing video games (especially RPGs and Rhythm Action Games) as I research them, keep thinking what is the best way for a player to co-create a living world with AI, toward their dream games.