# Learn to Unlearn for Deep Neural Networks: Minimizing Unlearning Interference with Gradient Projection

**Authors:** Tuan Hoang, Santu Rana, Sunil Gupta, Svetha Venkatesh  
**arXiv:** [2312.04095](https://arxiv.org/abs/2312.04095)  
**Date:** 2023-12-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2312.04095v1/

## Key points

- Recent data-privacy laws have led to interest in machine unlearning
- Machine unlearning is a process that removes the impact of specific training samples from a learned model
- The challenge is to discard this information without compromising knowledge about the remaining dataset efficiently
- A method called PGU, utilizing SGD for updating weights, has been adopted to address this challenge
- A study by Hoang et al. titled "Learn to Unlearn for Deep Neural Networks" explores machine unlearning further
- The study shows that PGU produces similar results to retraining models from scratch even when the training dataset is no longer accessible
- Experiments were conducted on forgetting classes using ResNet-18 and compared retrained and unlearnt models using various methods
- Results indicate that PGU effectively removes specific training samples while preserving knowledge about the remaining dataset

## Summary

Recent data-privacy laws have sparked interest in machine unlearning, a process that removes the impact of specific training samples from a learned model as if they were never part of the original dataset. The challenge lies in discarding this information without compromising knowledge about the remaining dataset, and doing so efficiently. To address this, a projected-gradient based method called PGU has been adopted. By utilizing SGD for updating weights, PGU can scale to any model and dataset size. A recent study by Hoang et al. titled "Learn to Unlearn for Deep Neural Networks" explores this concept further and provides evidence that their method produces similar results to retraining models from scratch even when the training dataset is no longer accessible. The study was accepted at WACV 2024 and conducted experiments on forgetting classes using ResNet-18 and compared retrained and unlearnt models using various methods. Results showed that PGU effectively removes specific training samples while preserving knowledge about the remaining dataset, making it a promising approach for addressing privacy concerns in machine learning models.