From learning models to revenue models: Professor Ningyuan Chen's balanced research approach to AI
When do we stop training AI models and start profiting from them?
This is the million (or billion) dollar question that Associate Professor Ningyuan Chen, of the University of Toronto Mississauga’s Institute for Management & Innovation and Department of Management, is tackling. With machine learning models—think self-driving cars and ChatGPT—Chen is focused on striking a balance between continually feeding these systems data and information and shifting the focus to revenue instead. Supported by an IMI Research Grant and in collaboration with Yufeng Zheng (PhD student, Rotman School of Management) and Rafid Mahmood (Assistant Professor, University of Ottawa), his work comes at a pivotal moment in the development of artificial intelligence (AI).
Chen’s current project focuses on the operations of companies that sell machine learning products, which are being used for everything from climate change modelling and health care diagnoses to traffic management and crime prevention. To create effective systems, companies train their AI models on large amounts of data. At some point however, these models need to become financially viable to be practical. Chen is interested in the intersection point of the two – the optimal point between data saturation and monetizing the model. IMI’s Research Grant has supported this work, giving Chen and his research assistants funds to present their work at internationally recognized conferences and support the cloud computing technology required to empirically test their work. At the root of his research is the question: how much data is enough data to make an intelligent AI model? Chen was inspired by the machine learning concept of neural scaling law, which describes how the performance of AI neural networks varies depending on the amount of data. In other words: the more data put into a large language model, the better it will perform. For companies selling these machine learning products, the dilemma of deciding to continue to feed data or monetize the model has been a common issue for AI developers - and where Chen’s curiosity lies.
Chen decided to approach the question through a combination of dynamic programming and online learning. The former is a computer programming technique which breaks down a larger problem into smaller sub-problems, saves them in memory, and then optimizes them to find the overall solution. In simpler terms, Chen describes it like a game of Tetris: “the player needs to decide the direction and rotation at any moment,” he says. “The decision impacts the state of the world, for instance, the configuration. This intertwined relationship between decision and state of the world captures the essence of dynamic decision-making.”
Online learning is a technique used in reinforcement learning, where the model learns from the environment while making decisions. “One can think of an amateur learning to invest in the stock market. You need to make decisions constantly, but you are learning the market at the same time,” says Chen. This technique prepares the AI model for various situations and unexpected circumstances.
Ultimately, Chen and his collaborators will propose guidelines for the ways in which these models learn scaling laws and independently decide whether to stop collecting new data simultaneously. Most of the challenges they’ve encountered have been theoretical (in easier terms: their proof worked successfully using empirical data), so in order to finalize the model, the team is verifying the proof theoretically and comparing it to the empirical results. To fully understand the magnitude of Chen’s work, consider the example of Tesla’s full self-driving cars. Tesla needs an enormous amount of data in order to train the self-driving AI model to be fully functional and anticipate any situation which might occur on the road. AI is trained on either simulated or real data. Simulated data imitates real data, and is less costly, but it doesn’t capture all potential scenarios, which are critical to creating reliable, fully self-driving vehicles. Real data is ideal in this kind of training; however, it’s incredibly costly and gathering data can be difficult.
Deciding what kind of data to use and when to stop training the model and begin monetizing are major decisions for companies who sell machine learning products and is still an under researched area of the field. “We believe the algorithm and framework we develop can help companies who provide AI services think about their strategic goals and think about the trade-off of model training and monetization in a better way,” says Chen. As AI continues to evolve, Chen’s work will have real impact, offering a new road map for companies navigating the line between innovation and implementation – ensuring that AI doesn’t just grow, but also pays off in the long term.
Dr. Ningyuan Chen is an Associate Professor at the University of Toronto Mississauga, in the Department of Management, cross-appointed to the Institute for Management & Innovation and the Rotman School of Management. He earned his Ph.D. from Columbia University and held positions at Yale and HKUST. His research focuses on machine learning and data-driven decision-making in business, with work published in top journals. He has received multiple research awards and funding, and collaborated with BAM, Polymatiks, and Alibaba on various projects. Learn more about Dr. Chen’s work here.
