Designing to Forget: Deep Semi-parametric Models for Unlearning

TL;DR

This paper introduces deep semi-parametric models that facilitate efficient unlearning of specific training samples without retraining, maintaining competitive performance in image tasks.

Contribution

The authors propose a novel family of semi-parametric models with a fusion module for explicit sample deletion, improving unlearning efficiency while preserving task accuracy.

Findings

SPMs achieve comparable performance to parametric models in image classification and generation.

SPMs reduce prediction gap by 11% on ImageNet compared to retrained baseline.

SPMs enable over 10x faster unlearning than existing parametric approaches.

Abstract

Recent advances in machine unlearning have focused on developing algorithms to remove specific training samples from a trained model. In contrast, we observe that not all models are equally easy to unlearn. Hence, we introduce a family of deep semi-parametric models (SPMs) that exhibit non-parametric behavior during unlearning. SPMs use a fusion module that aggregates information from each training sample, enabling explicit test-time deletion of selected samples without altering model parameters. Empirically, we demonstrate that SPMs achieve competitive task performance to parametric models in image classification and generation, while being significantly more efficient for unlearning. Notably, on ImageNet classification, SPMs reduce the prediction gap relative to a retrained (oracle) baseline by $11\%$ and achieve over $10\times$ faster unlearning compared to existing approaches on parametric models. The code is available at https://github.com/amberyzheng/spm_unlearning.