Reproduce “MultiDelete for Multimodal Machine Unlearning” 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

Jiali Cheng, Hadi Amiri

MultiDelete, a neuro-symbolic framework, generates high-fidelity synthetic data by guaranteeing 100% schema validity via Probabilistic Sentential Decision Diagrams and enhancing distributional coverage through convex optimization. This approach effectively decouples semantic reasoning from linguistic surface realization using a conditional Large Language Model, yielding data that is both logically consistent and diverse.

Key insights

  • Effective multimodal unlearning requires explicitly decoupling the learned associations between modalities for forgotten samples, rather than just removing knowledge of individual unimodal components.
  • Maintaining both the model's overall multimodal knowledge on retained data and its foundational unimodal representations of forgotten components is crucial for preserving the unlearned model's utility.
  • Traditional unimodal or modality-agnostic unlearning techniques are insufficient for multimodal contexts, often leading to incomplete forgetting or substantial performance degradation on retained data.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.