# Bayesian Inference Forgetting

**Authors:** Shaopeng Fu, Fengxiang He, Yue Xu, Dacheng Tao  
**arXiv:** [2101.06417](https://arxiv.org/abs/2101.06417)  
**Date:** 2021-01-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2101.06417v1/

## Key points

- The paper addresses the issue of enforcing the right to be forgotten in machine learning
- Existing approaches propose removing the influence of requested data points on learned models using influence functions
- Bayesian Inference Forgetting (BIF) framework extends forgetting algorithms to Bayesian inference
- BIF includes forgetting algorithms for variational inference and Markov chain Monte Carlo
- The proposed algorithms effectively remove the influence of individual data points on learned models
- Theoretical analysis demonstrates guaranteed generalizability of the proposed algorithms
- Experiments validate feasibility using Gaussian mixture models and Bayesian neural networks
- BIF framework offers effective forgetting algorithms for Bayesian inference, allowing compliance with individual requests without deleting entire models trained on massive resources.

## Summary

The paper titled "Bayesian Inference Forgetting" addresses the issue of enforcing the right to be forgotten in machine learning. This can lead to significant costs for companies. The right to be forgotten allows individuals to request the deletion of their personal data from models trained on massive datasets. Existing approaches propose removing the influence of requested data points on learned models using influence functions. However, these methods are not naturally well-defined in Bayesian inference. To overcome this limitation, the authors propose a framework called Bayesian Inference Forgetting (BIF) that extends the applicability of forgetting algorithms to Bayesian inference. The BIF framework includes forgetting algorithms for variational inference and Markov chain Monte Carlo. These algorithms can effectively remove the influence of individual data points on learned models. The paper provides theoretical analysis demonstrating that the proposed algorithms have guaranteed generalizability. Additionally, experiments conducted on Gaussian mixture models using synthetic datasets and Bayesian neural networks using the Fashion-MNIST dataset validate the feasibility of the proposed methods. Overall, this paper introduces a novel approach to address the challenges associated with enforcing the right to be forgotten in machine learning. The BIF framework offers effective forgetting algorithms for Bayesian inference, allowing companies to comply with individual requests without deleting entire models trained on massive resources.