As AI models evolve, we’ve seen how they harness the power of immense volumes of data. However, sometimes they tap into data they shouldn’t. When this happens, the system owner may be ordered by a regulatory agency or court to remove the data. Traditionally, when AI models needed to stop using data, the only reliable option was to rebuild the entire model from scratch. A new learning paradigm contributed by MSU researcher Sijia Liu allows AI models to selectively forget, or “unlearn,” specific data sets without the need to scrap the whole model.
Retraining an entire model can cost millions of dollars and take months, said Liu, Red Cedar Distinguished Associate Professor in the Department of Computer Science and Engineering. His machine unlearning approaches these challenges by enabling AI systems to responsibly update or remove unwanted knowledge.
In addition to providing an efficient way to comply with emerging regulations on privacy and copyrighted material, enabling AI systems to selectively unlearn what they already know reduces risks associated with outdated information or knowledge that could be used to do something harmful. As a result, when AI can unlearn what it shouldn’t have learned, it can become more trustworthy.
Liu’s unlearning technology has broad application potential for anyone building or relying on AI, from tech companies and government agencies to hospitals and financial institutions.
To learn more about Liu’s work:
MSU College of Engineering Media and Public Relations page