ContinualFlow: Learning and Unlearning with Neural Flow Matching

TL;DR

ContinualFlow is a novel framework that enables targeted unlearning in generative models using Flow Matching and energy-based reweighting, avoiding retraining and direct sample access.

Contribution

It introduces a principled unlearning method leveraging energy-based proxies and Flow Matching, with theoretical guarantees and practical validation.

Findings

Effective unlearning of undesired data regions

No need for retraining from scratch

Validated on 2D and image datasets

Abstract

We introduce ContinualFlow, a principled framework for targeted unlearning in generative models via Flow Matching. Our method leverages an energy-based reweighting loss to softly subtract undesired regions of the data distribution without retraining from scratch or requiring direct access to the samples to be unlearned. Instead, it relies on energy-based proxies to guide the unlearning process. We prove that this induces gradients equivalent to Flow Matching toward a soft mass-subtracted target, and validate the framework through experiments on 2D and image domains, supported by interpretable visualizations and quantitative evaluations.