# Online Continual Learning with Maximally Interfered Retrieval

**Authors:** Rahaf Aljundi, Lucas Caccia, Eugene Belilovsky, Massimo Caccia, Laurent Charlin, Tinne Tuytelaars  
**arXiv:** [1908.04742](https://arxiv.org/abs/1908.04742)  
**Date:** 2019-08-11  
**Summary page:** https://www.summarizepaper.com/en/paper/1908.04742v1/

## Key points

- Continual learning is a challenge for modern machine learning systems
- The online or "single-pass through the data" setting is a difficult setting to tackle
- Methods based on replay have shown to be effective approaches for continual learning and have matched or exceeded state-of-the-art benchmarks
- Randomly selecting samples from the replay memory or generative model is suboptimal
- Rahaf Aljundi et al. propose a controlled sampling of memories for replay by retrieving the samples that are most interfered with - those whose prediction will be most negatively impacted by foreseen parameter updates
- They show a formulation for this sampling criterion in both the generative replay and experience replay settings, resulting in consistent gains in performance and greatly reduced forgetting
- The authors' approach offers an improvement over existing methods
- Provides a promising direction for future research in continual learning

## Summary

Continual learning remains a significant challenge for modern machine learning systems. In particular, the online or "single-pass through the data" setting has recently gained attention as a natural and difficult setting to tackle. While methods based on replay, either generative or from stored memory, have shown to be effective approaches for continual learning and have matched or exceeded state-of-the-art benchmarks, they typically rely on randomly selecting samples from the replay memory or generative model which is suboptimal. Rahaf Aljundi et al. propose a controlled sampling of memories for replay by retrieving the samples that are most interfered with - those whose prediction will be most negatively impacted by foreseen parameter updates. They show a formulation for this sampling criterion in both the generative replay and experience replay settings, resulting in consistent gains in performance and greatly reduced forgetting. The authors' approach offers an improvement over existing methods and provides a promising direction for future research in continual learning.