AI Models Now Capable of 'Forgetting' Data: Breakthrough in Machine Unlearning

Scientists from the Tokyo University of Science (TUS) have devised a technique to allow large-scale AI models to selectively "forget" specific categories of data.

Advancements in AI have led to tools that could transform various fields, including healthcare and autonomous driving. But as the technology develops, so do the complexities and ethical challenges associated with it.

The trend of large-scale pre-trained AI systems, like OpenAI's ChatGPT and CLIP (Contrastive Language–Image Pre-training), has set new benchmarks for machine capabilities. These versatile models adept at handling numerous tasks consistently have become widely used in both professional and personal contexts.

However, this versatility comes at a significant cost. Maintaining and training these models demands substantial energy and time, raising sustainability concerns and necessitating advanced, costly hardware. Moreover, their generalist nature may limit their effectiveness when applied to specific functions.

As Associate Professor Go Irie, who spearheaded the research, points out, "In practical applications, classifying every type of object class is rarely necessary. For instance, an autonomous driving system only needs to recognize a few types of objects like cars, pedestrians, and traffic signs."

"Identifying things like food, furniture, or animal species is unnecessary. Keeping classes that aren't needed could reduce overall classification accuracy and waste computational resources, posing a risk for information leaks," he added.

A viable solution is training models to "forget" superfluous or irrelevant information, allowing them to concentrate solely on necessary tasks. Some current methods address this but operate on a "white-box" basis, requiring access to a model’s internal framework. Often, this is not possible for users.

More frequently used "black-box" AI systems, due to commercial and ethical limitations, hide their internal processes, making traditional forgetting methods impractical. To bridge this gap, the team turned to derivative-free optimization—avoiding reliance on a model's inaccessible inner workings.

Set for presentation at the upcoming Neural Information Processing Systems (NeurIPS) conference in 2024, the study introduces a technique called "black-box forgetting."

This involves adjusting input prompts (the instructions given to models) iteratively to progressively make the AI "forget" specific categories. Associate Professor Irie worked on this with Yusuke Kuwana and Yuta Goto from TUS, in partnership with Dr. Takashi Shibata from NEC Corporation.

They tested this approach using CLIP, a vision-language model skilled in image classification. The process relies on the Covariance Matrix Adaptation Evolution Strategy (CMA-ES), an evolutionary algorithm for refining solutions. Within the study, CMA-ES was used to assess and enhance prompts given to CLIP, eventually limiting its capacity to classify certain image categories.

As they advanced, the team encountered issues. Existing optimization methods struggled with scaling for a higher volume of targeted categories, prompting them to create a novel parametrization strategy known as "latent context sharing."

This new strategy divides latent context—a representation of information produced by prompts—into smaller, manageable bits. It assigns particular elements to a single token (word or character), while recycling others across various tokens, substantially simplifying the complexity of the problem. Importantly, this made the process usable for extensive forgetting tasks.

Through extensive testing on various image classification datasets, the researchers demonstrated the effectiveness of black-box forgetting—successfully making CLIP "forget" about 40% of target categories without accessing the AI model's internal framework.

This is the first documented success in driving selective forgetting within a black-box vision-language model, showing encouraging results.

Beyond its technical achievement, this innovation offers substantial potential for applications where precise task-specificity is essential.

Streamlining models for specific tasks could make them quicker, more resource-efficient, and capable of running on less advanced devices—accelerating the uptake of AI in areas previously regarded as impractical.

Another primary benefit is in image creation, where eliminating entire visual context categories could prevent models from generating unwanted or harmful content, such as offensive material or misinformation.

Crucially, this method tackles one of AI’s major ethical predicaments: privacy.

Large AI models are often trained on vast datasets that may contain unintended sensitive or outdated information. Addressing requests to remove such data is challenging, especially considering laws supporting the "Right to be Forgotten."

Completely retraining models to omit problematic data demands a significant investment of time and resources, yet neglecting it could have serious implications.

"Retraining a large model requires considerable energy," remarks Associate Professor Irie. "Selective forgetting, or machine unlearning, might offer an effective remedy for this issue."

This privacy-oriented application is particularly pertinent in critical sectors like healthcare and finance, where data sensitivity is of paramount importance.

As global efforts to propel AI progress, the Tokyo University of Science's black-box forgetting strategy offers an essential path—enhancing technological adaptability and efficiency while adding meaningful protections for end-users.

Even though the potential for misuse remains, solutions like selective forgetting illustrate a proactive approach from researchers towards addressing ethical and practical challenges.