Designing to Forget: Deep Semi-parametric Models for Unlearning
This work addresses the need for efficient machine unlearning, which is crucial for privacy and compliance in AI systems, by proposing a novel model design that improves unlearning speed and accuracy.
The paper tackles the problem of efficiently removing specific training samples from trained models by introducing deep semi-parametric models (SPMs) that enable explicit test-time deletion without altering parameters. It shows that SPMs achieve competitive performance in image classification and generation, reducing the prediction gap by 11% on ImageNet and offering over 10x faster unlearning compared to existing methods.
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.