ContinualFlow: Learning and Unlearning with Neural Flow Matching

2025-06-23

AI Summary

This work addresses the challenge of *targeted forgetting* in generative modeling—i.e., selectively removing specific regions of the data distribution without retraining or accessing the samples to be forgotten. We propose the first continual learning and selective forgetting framework that integrates flow matching with energy-based modeling. Methodologically, we design a proxy energy function to reweight the flow matching loss; theoretically, we prove this reweighting steers the learned vector field toward a *soft mass depletion* objective, enabling seamless, traceless forgetting. Our key contribution is the novel unification of flow matching and energy-based modeling for targeted distribution editing, accompanied by interpretable, visualization-enabled analysis. Experiments on 2D synthetic data and image benchmarks demonstrate that our method precisely excises target modes while preserving overall generation quality, significantly outperforming existing baselines.

Technology Category

• Machine Learning: Deep Generative Models & Autoencoders
• Computer Vision: Diffusion Models for Vision
• Natural Language Processing: Generation

Application Category

• User Modeling, Personalization and Recommendation: User modeling for targeted and personalized online advertising
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs

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.

Problem

• Targeted unlearning in generative models
• Subtract undesired data regions without retraining
• Energy-based proxies guide unlearning process

Innovation

• Energy-based reweighting loss for unlearning
• Flow Matching without retraining from scratch
• Energy-based proxies guide unlearning process