Black-box forgetting: A new method for tailoring large AI models

Training and operating large models consumes large amounts of energy and time, defeats sustainability goals, and limits the types of computers that can be deployed. Furthermore, in many real-world applications, people expect AI models to be able to fulfill specific roles rather than be a jack-of-all-trades. In this case, the model’s generalist capabilities may be useless or even counterproductive, reducing accuracy. Is there a way to utilize large-scale pre-trained models more efficiently by making them “forget” unnecessary information?

In a recently published paper Neural Information Processing System (NeurIPS 2024)A research team led by Go Irie, associate professor at Tokyo University of Science (TUS) in Japan, tried to solve this problem. They developed a method called “black-box forgetting,” whereby one can iteratively optimize the textual cues presented to a black-box visual language classifier model so that it selectively “forgets” what it can recognize. some categories. Co-authors of the study include Mr. Yusuke Kuwana and Mr. Yuta Goto from TusUS, and Dr. Takashi Shibata from NEC Corporation.

“In practical applications, it is rarely necessary to classify various object categories. For example, in autonomous driving systems, it is enough to recognize limited categories of objects (e.g., cars, pedestrians, traffic signs). We do not need to recognize food, furniture or animal species,” Dr. Irie explained. “Retaining categories that do not require identification may reduce overall classification accuracy and lead to operational disadvantages such as wasted computing resources and risk of information leakage.

Although some selective forgetting methods do exist in pre-trained models, these methods assume a white-box setup where the user has access to the internal parameters and architecture of the model. Users often encounter black boxes. For commercial or ethical reasons, they do not have access to the model itself or most of its information. Therefore, researchers must adopt a so-called derivative-free optimization strategy—one that does not require access to model gradients.

To this end, they extended a method called CMA-ES to use the image classifier model CLIP as the target model of this study. This evolutionary algorithm involves sampling various candidate cues to feed the model and evaluating the results through a predefined objective function, updating the multivariate distribution based on the calculated values.

However, for large-scale problems, the performance of derivative-free optimization techniques quickly deteriorates. As more and more categories need to be forgotten, the “potential context” used to optimize input prompts can grow to an unmanageable size. To solve this problem, the research team proposed a new parameterization technology called “potential situation sharing”. This approach involves decomposing the underlying context derived from the prompt into various smaller elements, which are considered to be “unique” to the prompt tag or “shared” among multiple tags. Optimizing these smaller units rather than large chunks of underlying context through optimization goals can significantly reduce the dimensionality of the problem, making it easier to handle.

The researchers validated their method using several benchmark image classification datasets, trying to make CLIP “forget” 40% of the categories in a given dataset. This marks the first study that aims to make pre-trained visual language models unable to recognize specific categories under black-box conditions, and the results are very promising based on a reasonable performance baseline.

This innovative approach has important implications in the fields of artificial intelligence and machine learning. It helps large models perform better at specialized tasks, extending their already astounding applicability. For example, another use is to prevent image generation models from generating unnecessary content by making them forget specific visual contexts.

Furthermore, the proposed method can help solve privacy issues, which is a growing concern in this field. “If the service provider is asked to remove certain information from the model, this can be done by removing the problematic samples from the training data and retraining the model from scratch. However, retraining large models consumes a lot of energy,” Said Dr. Irie, “Selective forgetting, or so-called machine forgetting, may provide an effective solution to this problem.” In other words, it could help develop solutions that protect the so-called “right to be forgotten,” which is a key factor in healthcare. and a particularly sensitive topic in finance.