October 6, 2023
Have you ever wondered how artificial intelligence (AI) learns from all the data it consumes? Well, it's not as simple as reading a book or watching a movie. AI uses complex mathematical models to process and understand huge amounts of information, such as text, images, audio, and video. These models are called large language models (LLMs), and they can do amazing things like write stories, answer questions, and generate images.
But what if some of the data that AI learns from is not supposed to be used? For example, what if the AI accidentally uses a song lyric, a movie quote, or a book passage that is protected by copyright? That means that the owner of that work has the legal right to control how it is used and distributed. If the AI uses it without permission, it could cause trouble for both the AI and the owner.
That's why some researchers from Microsoft have come up with a clever way to make AI forget what it shouldn't know. They used the famous Harry Potter books as an example. You know, the ones about the boy wizard who goes to a magical school and fights evil. The researchers wanted to see if they could erase all the knowledge of Harry Potter from an AI model, without affecting its other abilities.
How did they do it? They used a three-step technique that involved identifying, replacing, and fine-tuning. First, they trained another AI model to find all the words and phrases related to Harry Potter in the original model. Second, they replaced those words and phrases with generic ones that had nothing to do with Harry Potter. Third, they fine-tuned the original model on these new words and phrases, so that it would forget the old ones.
The result? The original model could no longer generate or recall anything about Harry Potter, even when given hints or prompts. It was as if Harry Potter never existed for the AI. The researchers claim that this is the first time that such a technique has been successfully applied to generative language models.
Why is this important? Well, it shows that AI can be adaptable and flexible, according to changing needs and situations. It also shows that AI can respect the rights and wishes of creators and owners of original works, by not using them without permission. And finally, it shows that fiction can be a powerful tool for teaching and testing AI, by providing clear and distinctive examples of what to learn and what to forget.
So next time you read a Harry Potter book or watch a Harry Potter movie, remember that you are not only enjoying a magical story, but also learning something that could help AI become smarter and more ethical. And who knows, maybe one day you will be able to chat with an AI that can create its own stories of wizards and magic!
For more info check out the source article: https://venturebeat.com/ai/researchers-turn-to-harry-potter-to-make-ai-forget-about-copyright-material/.