An Unlearning Framework for Continual Learning

Machine unlearning in continual learning requires a new approach: The paper identifies that existing unlearning methods fail when applied to continual learning (where models learn tasks sequentially), causing two critical problems: performance drops on retained tasks and "task relapse," where supposedly unlearned tasks reappear during future learning.

Data-free unlearning is essential for privacy: UnCLe addresses a key privacy concern by enabling unlearning without requiring access to original data. This is important because continual learning systems typically discard past data, making traditional unlearning methods impractical.

The proposed solution uses task-specific parameters: UnCLe employs a hypernetwork that generates task-specific network parameters based on task embeddings, then "unlearns" tasks by aligning these parameters with noise—allowing the model to forget information without destabilizing previously learned knowledge.

Minimal disruption to retained knowledge is achievable: Empirical testing on vision datasets demonstrates that UnCLe can perform multiple sequential learning and unlearning operations while preserving performance on tasks that should be retained, addressing a major challenge in responsible AI deployment.

Practical implications for AI safety and compliance: This framework enables AI systems to respond to safety concerns and privacy requests (like GDPR "right to be forgotten" requests) in real-time, even as models continue learning new tasks—a critical capability for responsible AI systems in production environments.