Tokyo University of Science (TUS) researchers have created a technique that allows large-scale AI models to “forget” particular data classes on a selective basis.
Advances in AI have produced technologies that have the potential to transform a number of industries, including autonomous driving and healthcare. But as technology develops, so do its intricacies and moral dilemmas.
Expectations for machines have changed as a result of the paradigm shift toward large-scale pre-trained AI systems, such CLIP (Contrastive Language–Image Pre-training) and ChatGPT from OpenAI. These extremely generalist models have been widely used for both personal and professional purposes since they can consistently perform a wide range of activities with accuracy.
But this adaptability comes at a high cost. In addition to requiring state-of-the-art hardware that is far more costly than typical computers, training and operating these models requires enormous amounts of time and energy, posing sustainability problems. These problems are made worse by the possibility that generalist inclinations could reduce AI models' effectiveness when used for particular tasks.
According to Associate Professor Go Irie, who oversaw the study, “in practical applications, the classification of all kinds of object classes is rarely required.” For instance, it would be adequate for an autonomous driving system to recognize a small number of object classes, such as automobiles, people, and traffic signs.
We wouldn't have to identify different kinds of food, furnishings, or animals. In addition to operational drawbacks like computing resource waste and information leakage risk, maintaining classes that are not required to be recognized may reduce overall classification accuracy.
Training models to “forget” extraneous or useless information could be a solution, simplifying their operations to concentrate only on what is important. Although some current approaches currently address this need, they frequently operate on the assumption that users have access to the core architecture and parameters of a model in a “white-box” fashion. However, users frequently do not receive this kind of visibility.
Due to ethical and financial constraints, “black-box” AI systems are more prevalent and hide their internal workings, making conventional forgetting methods unfeasible. The research team used derivative-free optimization, which avoids depending on a model's inaccessible internal workings, to close this gap.
Advancing through forgetting
The work presents a mechanism called “black-box forgetting,” and it is scheduled to be presented at the 2024 Neural Information Processing Systems (NeurIPS) conference.
In iterative rounds, the procedure alters the input prompts (text instructions sent to models) so that the AI gradually “forgets” particular classes. Co-authors Yusuke Kuwana and Yuta Goto, both from TUS, and Dr. Takashi Shibata from NEC Corporation worked with Associate Professor Irie on the project.
The researchers focused their tests on CLIP, a vision-language model that can classify images. The Covariance Matrix Adaptation Evolution Strategy (CMA-ES), an evolutionary algorithm intended to optimize solutions gradually, is the foundation of the approach they created. In this work, CLIP's ability to categorize particular image categories was suppressed by using CMA-ES to assess and refine the prompts that were given to it.
As the project developed, difficulties surfaced. The researchers came up with a novel parametrisation mechanism called “latent context sharing” because existing optimization techniques had trouble scaling up for higher volumes of targeted categories.
This method divides latent context, which is a representation of data produced by prompts, into smaller, easier-to-manage chunks. They significantly decreased the complexity of the problem by assigning some aspects to a single token (word or letter) and reusing others over several tokens. Importantly, even for large-scale forgetting applications, this made the procedure computationally tractable.
The researchers confirmed the effectiveness of black-box forgetting—making CLIP “forget” about 40% of target classes without direct access to the internal architecture of the AI model—through benchmark tests on many image classification datasets.
With encouraging results, this study is the first successful attempt to induce selective forgetting in a black-box vision-language model.
Benefits of helping AI models forget data
Beyond its technical inventiveness, this invention has a lot of promise for practical uses where accuracy is crucial for a given task.
The adoption of AI in previously unthinkable fields might be accelerated by simplifying models for specialized tasks to make them faster, more resource-efficient, and able to operate on less powerful machines.
Another important application is in picture generation, where models may be able to avoid unintentionally producing inappropriate or damaging content, such as false information, by forgetting entire categories of visual context.
Most significantly, this approach tackles privacy, one of AI's biggest ethical conundrums.
AI models, particularly large-scale ones, are often trained on massive datasets that may inadvertently contain sensitive or outdated information. Requests to remove such data—especially in light of laws advocating for the “Right to be Forgotten”—pose significant challenges.
Retraining entire models to exclude problematic data is costly and time-intensive, yet the risks of leaving it unaddressed can have far-reaching consequences.
“Retraining a large-scale model consumes enormous amounts of energy,” notes Associate Professor Irie. “‘Selective forgetting,’ or so-called machine unlearning, may provide an efficient solution to this problem.”
These privacy-focused applications are especially relevant in high-stakes industries like healthcare and finance, where sensitive data is central to operations.
The black-box forgetting technique from Tokyo University of Science paves a crucial way ahead as the race to develop AI quickens globally—not only by increasing the technology's efficiency and adaptability but also by introducing critical user safeguards.
Even though there is still room for abuse, techniques like selective forgetting show that scientists are actively tackling moral and practical issues.
