Huang uses machine unlearning to make AI trustworthy

What if AI learned it wrong? Training artificial intelligence models means taking in massive data sets to establish the model’s parameters. But the challenge with opening the information gate so wide is that undesirable, incorrect or copyright-protected information may flow through the pipeline as well.

Jun Huang, an assistant professor in electrical engineering and computer science at South Dakota State University, is working with a doctoral student to refine the relatively new field of machine unlearning.

Huang, who joined the SDSU faculty in 2023 after serving at Baylor University, said his academic roots were in wireless communication networks, but now a large part of his research focus has shifted to AI. His first professional paper on the topic was published in 2022, when he held a non-tenure-track position at Baylor University in Waco, Texas.

Huang and graduate student ZiHao Ding began studying machine unlearning at the start of the 2025-26 school year.

During that time, they have had three top-tier papers selected for publication and his proposal “Trustworthy Federated Unlearning for Mobile Autonomous Systems” was funded by the South Dakota Board of Regents for $90,000 from July 31, 2026, to August 31, 2027. The funding allowed him to conduct research with two robotic arms, four ground rovers, and three Jetson boxes that are used in large language models.

Huang hopes the grant serves as a starter for additional external funding as the work, while promising, is just at the initial stage.

More than forgetting — unlearning

“We are trying to make artificial intelligence more useful in daily life, not just in big data centers. We are looking at how AI in cars, drones, phones, and small devices can retain task-critical knowledge over time rather than forgetting what they have learned. We are also exploring ways for these devices to learn from each other when the internet is unreliable or computing power is limited.”

Another part of their work is about helping AI remove knowledge that is private, outdated or wrong on purpose.

It is the latter part on which the Board of Regents’ grant focuses upon. 

Huang emphasized the difference between unlearning and forgetting. In human terms, we’re likely to forget what we learn in high school biology or chemistry because our brain has taken in so much other information since then. In computing, forgetting is usually an unintended consequence of AI models overwriting previously learned information.

But that doesn’t mean an individual’s private data or an agency’s sensitive information doesn’t still exist within that model.

    Unlearning is an intentional process. Its goal is to remove the influence of designated information while preserving the model’s useful knowledge and performance.

    Without an effective unlearning method, developers may need to delete the affected training data and retrain the complete model from scratch. Huang said this process can require substantial time, computing power and expense.

Follow-up NSF is being sought

At this point, they are using models with known defects. In one example, they are using a dirty lens on a camera that is recording plant growth. In another simple case, a robotic arm is trained to pick up a red coffee cup rather than a black one. 

Huang and Ding have not yet taught an AI model to detect between flawed and accurate images, but they are seeking a National Science Foundation grant to do so.

“Right now, we feel this can have a wide application,” including the training of mobile autonomous terminals such as self-driving cars, Huang said.

In addition to his research, he teaches upper-level and graduate courses in computer networks, operating systems, mobile and cloud computing, and distributed machine learning systems.