HomeWinBuzzer NewsMicrosoft Researchers Develop Unlearning Technique for Large Language Models

Microsoft Researchers Develop Unlearning Technique for Large Language Models

Microsoft's new un-training method modifies models to omit specific knowledge without a complete retraining.


As the discussion heightens surrounding copyrighted materials in the training of large language models (LLMs) like OpenAI ChatGPT, Meta's , and Anthropic's Claude 2, a new approach is proposed by researchers Ronen Eldan of Microsoft Research and Mark Russinovich of Microsoft Azure. In a paper published on the non-peer reviewed site arXiv, they suggest a method to amend these models to forget certain knowledge, without requiring a total restructure or retraining.

Their experiment used Meta's Llama 2-7B model and managed to erase all knowledge of the Harry Potter books within it, including characters and plotlines. Originally, the model required over 184,000 GPU-hours to train, but the alterations only consumed about 1GPU hour – demonstrating the swift efficiency of this new technique.

Three-Step Technique to “Unlearn” Information

researchers accomplished this innovation using a three-part scheme to selectively remove specific information from the models. They began with training a model on the target data, in this case, the Harry Potter series. Following this, they replaced unique phrases from the series with generic alternatives, making predictions based on a model lacking that specific training. Finally, the baseline model was fine-tuned based on these altered predictions, resulting in the removal of the original material from the model's memory.

The authors believe their method to be the first of its kind to effectively encourage LLMs to forget, stating: “To the best of our knowledge, this is the first paper to present an effective technique for unlearning in generative language models.”

Effectiveness and Future Applications

The effectiveness of this unlearning process was evaluated by assessing the model's ability to generate or discuss content related to Harry Potter. Before the finetuning process, the model could easily produce intricate details from the series. Afterward, however, the model appeared to have “forgotten” the majority of those narratives. Importantly, this method did not affect the model's performance on conventional benchmarks such as ARC, BoolQ, and Winogrande.

This proof-of-concept provides a crucial step toward creating larger language models that are more adaptable, responsible, and legally compliant for future use. It allows for the potential modification of AIs to suit specific organizational, user requirements, and societal values. There is a need for further research, but this technique signifies a promising direction for the development of AI models.

However, the authors noted that further testing was necessary and that their technique might be more suitable for fictional narratives. They concluded by stressing the importance of refining their method and the need for additional research to ensure that AI systems could adapt dynamically to shifting priorities over time.

Copyright Battlefield Over AI Training

A lawsuit has been filed against OpenAI by a group of writers, including Michael Chabon and David Henry Hwang, who are well-known in the literary world. They claim that used their works without permission to train its AI model, . They also sued Meta Inc. for doing the same thing.

This is not the first time that OpenAI and Meta have been accused of such actions. Earlier this year, Sarah Silverman, Christopher Golden, and Richard Kadrey said that the companies or their affiliates illegally downloaded their books from websites that offer pirated books for free, such as Bibliotik, Library Genesis, Z-Library, and others. They said that their books were among those that were stolen and used to train the AI models.

Paul Tremblay and Mona Awad also took legal action against OpenAI in June for similar reasons. The current lawsuit not only asks for monetary damages but also wants the court to stop OpenAI from engaging in what they call “unlawful and unfair business practices.”

In July, some of the leading news publishers also thought about suing AI companies for violating their copyrights. They said that the AI firms were harming their intellectual property rights and their business model by copying, summarizing, or rewriting their articles and spreading them on different platforms, such as websites, apps, or social media.

The latest lawsuit against filed last month in federal court in New York, was organized by the Authors Guild and includes 17 authors, including John GrishamJodi Picoult, and Game of Thrones creator George R.R. Martin.

Luke Jones
Luke Jones
Luke has been writing about all things tech for more than five years. He is following Microsoft closely to bring you the latest news about Windows, Office, Azure, Skype, HoloLens and all the rest of their products.

Recent News