CLEAR

less than 1 minute read

🧠 CLEAR: first multimodal benchmark to make models forget what we want them to forget

With privacy concerns rising, we sometimes need our models to β€œforget” specific information - like a person’s data - while keeping everything else intact. Researchers just released CLEAR, the first benchmark to test how well this works with both text and images.

❌ Bad news: Current methods either fail to truly forget or end up forgetting way too much. It’s like trying to remove a single ingredient from a baked cake!

✨ But there’s hope: Adding simple mathematical constraints (L1 regularization) during the forgetting process significantly improves results.

🎯 Key insights:

βœ… The benchmark tests forgetting on 200 fictional personas

β€£ 3,770 visual Q&A pairs

β€£ 4,000 textual Q&A pairs

β€£ Additional real-world tests

πŸ›‘ Most current forgetting methods don’t work well with both text and images

β€£ They either remember what they should forget

β€£ Or they forget too much unrelated information

✨ Simple mathematical constraints work surprisingly well

β€£ L1 regularization prevents excessive forgetting

β€£ Works especially well with the LLMU method

πŸ‘‰ Read the full paper here: https://huggingface.co/papers/2410.18057

image.png
image.png