Reproduce “Zero-Shot Class Unlearning in CLIP with Synthetic Samples” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
A. Kravets, V. Namboodiri
The University of Bath research introduces "Lip," a zero-shot method for class unlearning in CLIP that generates synthetic samples and applies a multimodal Lipschitz regularization. This approach effectively removes specific class knowledge from both visual and textual encoders while maintaining performance on unrelated classes and datasets.
Key insights
- Truly zero-shot class unlearning in large multimodal models like CLIP is achievable by generating effective synthetic forget samples.
- Forgetting in CLIP requires joint consideration of both visual and textual encoders, as solely updating one modality is insufficient.
- A controlled, iterative approach with selective parameter updates is crucial for efficient unlearning, preventing widespread degradation of model utility.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.