Researchers from the University of Washington, along with colleagues from Princeton, the University of Chicago, USC, and Google, conducted a study on unlearning techniques used to improve the performance of generative AI models. These techniques are aimed at making models forget specific information that they have picked up from training data, such as sensitive private data or copyrighted material.
However, the study found that the current unlearning methods available may actually degrade the models, making them less capable of answering basic questions. Weijia Shi, a researcher on the study, highlighted that there are currently no efficient methods that enable a model to forget specific data without a significant loss of utility.
Generative AI models learn by analyzing patterns in large datasets, such as text, images, or videos. For example, a model trained to autocomplete messages may suggest the phrase “Looking forward to hearing back” based on patterns in the data it has been fed. These models, like GPT-4o, are usually trained on publicly available data without the consent of the data owners, leading to copyright concerns.
The issue of copyright has led to increased interest in unlearning techniques, which could help remove sensitive or copyrighted information from models. Unlearning algorithms work by steering models away from specific data, influencing their predictions to avoid outputting certain information. However, the study revealed that while these algorithms can make models forget certain information, they also impact the models’ general question-answering abilities.
The researchers developed a benchmark called MUSE to evaluate the effectiveness of unlearning algorithms. The benchmark tests whether a model can forget specific data, such as excerpts from Harry Potter books, while still retaining general knowledge related to the text. The study showed that existing unlearning methods face challenges in disentangling knowledge within the model, leading to trade-offs in performance.
Shi emphasized the need for further research to address the limitations of current unlearning techniques. While vendors may see unlearning as a potential solution to data privacy concerns, technical breakthroughs are needed to make unlearning more effective. In the meantime, vendors will need to explore alternative strategies to ensure their models do not inadvertently reveal sensitive or copyrighted information.
