Reproduce “Stochastic Meta-Unlearning: Bridging Language Backbone and Multimodal Unlearning” on OpenResearch

About this paper

Zijie Liu, Jinhao Duan, Gaowen Liu, Sijia Liu, Tianlong Chen

Stochastic Meta-Unlearning (SMU) introduces a bilevel meta-optimization framework to bridge the gap between localized language backbone unlearning and comprehensive multimodal forgetting in Vision-Language Models (VLMs). The framework consistently achieved an average of 9.16% Forget accuracy, while maintaining 75.23% Retain accuracy and 68.45% Test accuracy, outperforming direct VLM unlearning baselines.

Key insights

• Applying unlearning solely to a VLM's language backbone with text-only feedback is insufficient to guarantee reliable forgetting in a full multimodal context, highlighting the "multimodal recovery" phenomenon.
• A bilevel meta-optimization approach, where inner-loop updates are text-side and outer-loop feedback is VLM-level, effectively bridges the component-system mismatch, enabling localized unlearning to achieve system-wide multimodal forgetting.
• Meta-learning an "unlearning-ready" initialization enhances the transferability of unlearning capabilities to new forgetting targets and different unlearning operators, making the method more robust and adaptable.

Summary by alphaXiv.