Researchers at The University of Texas at Austin have pioneered a groundbreaking approach to address the challenges faced by artificial intelligence encountering copyright-protected and violent content. This innovative method, known as "machine unlearning," is specifically tailored for image-based generative AI and aims to block and eliminate undesirable content without the need to start the training process from scratch.
Professor Radu Marculescu from the Cockrell School of Engineering emphasized the significance of this advancement, stating, "When you train these models on such massive data sets, you’re bound to include some data that is undesirable." The conventional method of rectifying problematic content involved discarding all existing data and retraining the model entirely. However, with the introduction of machine unlearning, this process can now be accomplished without the arduous task of starting over.
The utilization of generative AI models, predominantly trained using vast internet data, has highlighted the prevalence of copyrighted material, personal data, and inappropriate content. This issue was underscored by The New York Times' legal action against OpenAI for allegedly using its articles without permission for training purposes.
Guihong Li, a graduate research assistant involved in the project, stressed the necessity of integrating mechanisms to prevent copyright infringement and the dissemination of harmful content in generative AI models intended for commercial applications.
The research primarily focuses on image-to-image models that manipulate input images based on contextual instructions. The newly developed machine unlearning algorithm grants the capability to eliminate flagged content without necessitating a complete model retraining. Human moderation teams are involved in overseeing and removing undesirable content, providing an additional layer of verification and responsiveness to user feedback.
Machine unlearning, a burgeoning field primarily applied to classification models, is relatively unexplored in the context of generative AI, especially concerning images. The research team, including Hsiang Hsu and Chun-Fu (Richard) Chen from JPMorgan Chase's Global Technology Applied Research group, will present their findings at the upcoming International Conference on Learning Representations in Vienna.
This innovative approach marks a significant step towards ensuring the ethical and lawful use of generative AI models, safeguarding against copyright infringement and the propagation of harmful content.