The performance of large-scale pre-trained AI models has significantly improved lately, a trend highlighted by models like CLIP and ChatGPT, which are adept at handling a variety of tasks. This versatility has led to their increased adoption, but it also presents certain challenges.
Training and running these expansive models require substantial amounts of time and energy, raising concerns about sustainability and restricting the types of hardware they can be used on. Additionally, users often need AI models for specific tasks rather than as generalists, meaning the broad capabilities of these models can sometimes hinder performance and decrease accuracy. Is there a way to make these large-scale pre-trained models more effective by having them ‘forget’ irrelevant data?
A research team, led by Associate Professor Go Irie from the Tokyo University of Science (TUS), Japan, has recently tackled this issue in a paper set to be presented at Neural Information Processing Systems (NeurIPS 2024). They introduced a technique known as “black-box forgetting,” which allows users to gradually modify the text prompts given to a black-box vision-language classifier to selectively forget certain recognizable classes. The study also involved co-authors Mr. Yusuke Kuwana and Mr. Yuta Goto from TUS and Dr. Takashi Shibata from NEC Corporation.
“In many applications, it isn’t necessary to classify all types of object classes,” Dr. Irie notes. “For instance, in autonomous driving, recognizing just a few classes like cars, pedestrians, and traffic signs is sufficient. Identifying unrelated categories like food or furniture isn’t needed and can actually reduce classification accuracy, waste computational resources, and pose risks for data leakage.”
While some selective forgetting methods exist for pre-trained models, they typically require a white-box scenario where users can access the model’s internal parameters and structure. However, users often work with black-box systems and cannot access these details for various commercial and ethical reasons. Therefore, the researchers needed to apply a derivative-free optimization approach—one that does not depend on having access to the model’s gradients.
To achieve this, they adapted a method called CMA-ES and used the CLIP image classifier as their experimental model. This evolutionary algorithm works by testing various candidate prompts with the model and evaluating the outcomes against predetermined objectives, and then updating a multivariate distribution based on these evaluations.
Nonetheless, the efficiency of derivative-free optimization strategies declines rapidly with larger problems. As more classes are targeted for forgetting, the ‘latent context’ used for optimizing the input prompts becomes increasingly complex. To overcome this, the research team created a new parameterization technique called ‘latent context sharing.’ This technique breaks down the latent context generated from prompts into smaller, manageable elements, identifying some as ‘unique’ to a token and others as ‘shared’ across multiple tokens. Focusing on these smaller segments makes the optimization challenge much less complex.
The researchers tested their technique with various benchmark image classification datasets, attempting to make CLIP ‘forget’ about 40% of the classes within those datasets. This study represents a significant advancement as it aims to cause a pre-trained vision-language model to not recognize specific classes even under black-box conditions, and the early results are quite promising.
This innovative approach has noteworthy implications in artificial intelligence and machine learning, potentially enhancing the performance of large-scale models in specialized roles, thereby widening their already impressive applications. Another possible use is preventing image generation models from creating inappropriate content by instructing them to forget certain visual contexts.
Additionally, this method might address rising privacy concerns in the industry. “When a service provider needs to remove specific information from a model, it often requires retraining the model from scratch and eliminating the problematic data from the training set, which is incredibly energy-intensive,” Dr. Irie explains. “Selective forgetting, also known as machine unlearning, could offer a more efficient alternative.” This means it could contribute to solutions that support the “Right to be Forgotten,” a particularly delicate issue in sectors like healthcare and finance.
