Pull-to-Outlier \& Contrastive Objective-level (POCO) Unlearning: A Framework for Sample and Objective Forgetting

TMLR Paper6622 Authors

Abstract: Current Machine Unlearning (MU) methods require full retraining or extensive fine-tuning, lack formal removal criteria, and focus only on sample-level forgetting, limiting their practicality. We address these gaps with two lightweight, projection-only techniques operating above frozen feature extractors. Pull-to-Outlier Unlearning (POU) offers a transparent, unsupervised geometric removal method by displacing embeddings of unwanted samples or entire classes into synthetic outlier regions, while preserving downstream performance and distilling knowledge of the remaining data. To the best of our knowledge, Contrastive Objective-level Unlearning (COU) is the first method to remove learned objectives. It perturbs projection weights to eliminate a target task’s influence. Then it realigns the original data manifold, which can provide the possibility for managing agentic learning behaviors. We validate POU on CIFAR10, CIFAR100, and Caltech-256 with ResNet-based backbones, showing efficient instance and class forgetting with minimal impact on retained accuracy. COU is tested on DINO and CLIP feature representations, demonstrating effective objective-level erasure while preserving all non-target tasks.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Rahaf_Aljundi1

Submission Number: 6622