Reproduce “ContinualFlow: Learning and Unlearning with Neural Flow Matching” on OpenResearch

About this paper

Lorenzo Simone, Davide Bacciu, Shuangge Ma

ContinualFlow presents a framework for targeted unlearning in generative models by modulating flow trajectories through energy-based reweighting, enabling effective suppression of undesired content even without direct access to the 'forget set'. This approach achieves unlearning performance comparable to retraining from scratch, while offering greater efficiency and demonstrating the novel capability of reversible unlearning.

Key insights

• Energy functions can provide a principled, continuous, and differentiable mechanism to steer generative trajectories, offering fine-grained control over unlearning without direct forget set access.
• The framework enables modular and composable control over unlearning dynamics, allowing for adaptation to evolving or combined unlearning criteria.
• ContinualFlow demonstrates the invertibility of energy functions, allowing the recovery of previously unlearned content by reversing the energy guidance, suggesting a new paradigm for reversible and privacy-aware generative content control.

Summary by alphaXiv.