Machine learning unlearning techniques have emerged as an important way to scrub generative AI models of unwanted elements, such as sensitive personal data or protected content, that may have been inadvertently absorbed during the training phase. However, these methods have significant drawbacks. A recent collaborative study involving experts from the University of Washington, Princeton University, University of Chicago, USC, and Google highlights a troubling trade-off: while these techniques strive to eliminate irrelevant data, they can severely impair the AI's fundamental cognitive functions.
The findings reveal that currently popular delearning techniques can render advanced models like OpenAI's GPT-4 and Meta's Llama 3.1 405B substantially less capable of processing even basic queries, rendering them virtually ineffective.
What is Unlearning in Machine Learning?
Machine unlearning is a relatively new concept in the field of artificial intelligence, specifically with regard to large-scale language models (LLMs). Simply put, machine unlearning is the process of making a machine learning model forget specific data it previously learned from. This becomes important when that data contains sensitive personal information or copyrighted material that should not have been included in the training set in the first place.
For non-tech savvy peopleMachine learning, the foundation of artificial intelligence, trains computers to interpret data and make decisions. There are three main types of machine learning: supervised learning, unsupervised learning, and reinforcement learning.
Supervised learning You use labeled data (examples with known outcomes) to predictively train your model. This method is similar to learning from an answer key and works best when:
- Classification tasksSuch as identifying whether an email is spam or not.
- Regression TasksReal estate price predictions, etc.
Unsupervised learning It works without labeled data, allowing the model to identify patterns and structure on its own. It is similar to self-learning without explicit guidance and is useful for:
- ClusteringModels group similar data points together, such as customer segmentation.
- AssociationThis involves finding commonalities in data, such as market basket analysis, where customers who buy one product also buy another product.
Reinforcement learning It learns through trial and error and uses rewards and penalties to shape the agent's behavior in the decision-making process, much like a trainer uses treats to teach a dog new tricks:
- Games and SimulationsThis is where agents learn winning strategies.
- Robot MovementSuitable for tasks that require a precise sequence of actions.
Each learning type utilizes a unique approach to digesting and processing information, selected based on the specific requirements of the task and the availability of data.
The challenge of unlearning
Language models are trained using huge amounts of text data collected from various sources. This data may unintentionally contain personal information or copyrighted content. If a data owner (an individual or entity that owns the rights to a dataset) identifies their data in the model and wants it removed, for example due to privacy concerns or copyright infringement, the ideal solution is to simply remove this data from the model.
But it's not easy to completely remove specific data from a language model that's already been trained from billions of other data points. This process, often called “retraining,” adjusts the model as if the specific data had never been part of the learning process in the first place. This is typically “intractable,” or impractical, for modern large-scale models due to the complexity and sheer amount of data to process.
AI and Machine Learning Trends to Watch in 2024
Approximate Machine Learning Unlearning Algorithms
The difficulty of accurate unlearning has led researchers to develop several “approximate unlearning algorithms.” These are methods designed to remove the influence of unwanted data from a model without having to rebuild the model from scratch. However, evaluating the effectiveness of these algorithms can be difficult. Historically, evaluations have been limited, and it has not been well understood whether these algorithms successfully meet the needs of both data owners (who want to forget the data) and model deployers (who want to maintain the validity of their models).
Introducing MUSE
To address these evaluation challenges, this work proposes MUSE, a comprehensive benchmark for evaluating machine unlearning. MUSE tests unlearning algorithms against six criteria that are considered desirable characteristics for models that undergo unlearning.
- No need for literal memorization: Models do not need to memorize exact phrases or sentences.
- No need to memorize knowledge: Detailed knowledge gained from specific data should not be retained.
- No privacy leaks: Your personal information will not be leaked.
- Save Utility: The model should work fine with other data that is not subject to deletion.
- Scalability: Large scale and multiple data deletion requests need to be handled efficiently.
- sustainability: Consecutive unlearning requests must be managed without performance degradation.
How do I untrain a model?
Generative AI models operate without what we think of as true intelligence. Rather, these systems work based on statistical analysis to predict patterns across a wide range of data, from textual content and images to audio and video, by processing large numbers of examples, including videos, audio recordings, essays, etc. For example, when presented with the phrase “We look forward to hearing from you…”, a model trained to autocomplete emails might predictively end with “We look forward to hearing from you…” based solely on the repetition it has observed in the data, without any consideration of human predictions.
Primarily, these models, including the advanced GPT-4o, get their training from publicly available websites and datasets under the banner of “fair use.” This practice, which is defended by its developers, involves scraping this data without the consent, compensation, or approval of the original data owners, and has led to legal challenges from various copyright holders calling for reform.
Against this backdrop, the concept of unlearning machine learning is gaining attention. Recently, in collaboration with academic partners, Google launched a competition aimed at stimulating the development of new unlearning techniques that would make it easier to remove sensitive content, such as medical records or sensitive images, from AI models upon request or legal demand. Historically, these models have often inadvertently captured personal information, ranging from phone numbers to sensitive data, due to the way they were trained. While some companies have put mechanisms in place to allow data to be excluded from future training, these do not apply to models already in use, positioning unlearning as a more comprehensive solution to data removal.
However, unlearning a machine is not as simple as simply deleting a folder: today's unlearning techniques employ sophisticated algorithms designed to redirect models away from unnecessary data, including by subtly adjusting the model's prediction mechanisms so that a given piece of data is never, or only very rarely, returned.
In this study, we applied these criteria to evaluate popular unlearning algorithms on a language model trained with 7 billion parameters, using datasets including Harry Potter books and news articles. Results show that while most algorithms were able to prevent the model from verbatim memorization or knowledge retention to some extent, only one algorithm was able to do so without causing significant privacy leakage. Furthermore, these algorithms generally fell short in maintaining the overall usefulness of the model, especially when dealing with large or multiple unlearning requests.
The findings of this study highlight a critical gap in the practical application of unlearning algorithms: they often fall short of the standards necessary for effective and secure data deletion. This has important implications for privacy advocates and AI developers alike.
In summary, machine learning unlearning is a promising area for addressing important ethical concerns in AI development, but much work remains to be done to make these techniques practical and trustworthy. The MUSE benchmark aims to aid this development by providing a robust framework for evaluating and improving unlearning algorithms.
Image credit: Kerem Gülen/Midjourney
