CLEAR
π§ CLEAR: first multimodal benchmark to make models forget what we want them to forget
With privacy concerns rising, we sometimes need our models to βforgetβ specific information - like a personβs data - while keeping everything else intact. Researchers just released CLEAR, the first benchmark to test how well this works with both text and images.
β Bad news: Current methods either fail to truly forget or end up forgetting way too much. Itβs like trying to remove a single ingredient from a baked cake!
β¨ But thereβs hope: Adding simple mathematical constraints (L1 regularization) during the forgetting process significantly improves results.
π― Key insights:
β 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
π Most current forgetting methods donβt work well with both text and images
β£ They either remember what they should forget
β£ Or they forget too much unrelated information
β¨ Simple mathematical constraints work surprisingly well
β£ L1 regularization prevents excessive forgetting
β£ Works especially well with the LLMU method
π Read the full paper here: https://huggingface.co/papers/2410.18057
